ercim-news.ercim.eu Number 125 April 2021 ERCIM NEWS

Special theme: Brain-inspired Computing

Also in this issue Research and Innovation: Human-like AI JoinCt ONTENTS

Editorial Information JOINT ERCIM ACTIONS

ERCIM News is the magazine of ERCIM. Published quarterly, it reports 4 ERCIM-JST Joint Symposium on Big Data and on joint actions of the ERCIM partners, and aims to reflect the contribu - Artificial Intelligence tion made by ERCIM to the European Community in Information Technology and Applied Mathematics. Through short articles and news 5 ERCIM “Alain Bensoussan” Fellowship Programme items, it provides a forum for the exchange of information between the institutes and also with the wider scientific community. This issue has a circulation of about 6,000 printed copies and is also available online. SPECIAL THEME

ERCIM News is published by ERCIM EEIG The special theme “Brain-inspired Computing” has been BP 93, F-06902 Sophia Antipolis Cedex, France coordinated by the guest editors Robert Haas (IBM +33 4 9238 5010, [email protected] Research Europe) and Michael Pfeiffer (Bosch Center for Director: Philipp Hoschka, ISSN 0926-4981 Artificial Intelligence)

Contributions Introduction to the Special Theme Contributions should be submitted to the local editor of your country 6 Brain-inspired Computing by Robert Haas (IBM Research Europe) and Michael Copyrightnotice Pfeiffer (Bosch Center for Artificial Intelligence) All authors, as identified in each article, retain copyright of their work. ERCIM News is licensed under a Creative Commons Attribution 4.0 8 Biologically Inspired Dynamics Enhance Deep International License (CC-BY). Learning by Stanisław Woźniak, Thomas Bohnstingl and Advertising Angeliki Pantazi (IBM Research – Europe) For current advertising rates and conditions, see https://ercim-news.ercim.eu/ or contact [email protected] 9 Effective and Efficient Spiking Recurrent Neural Networks ERCIMNewsonlineedition: ercim-news.ercim.eu/ by Bojian Yin (CWI), Federico Corradi (IMEC Holst Centre) and Sander Bohté (CWI) Nextissue: July 2021: Privacy-preserving Computation 11 Back-propagation Now Works in Spiking Neural Networks! Subscription by Timothée Masquelier (UMR5549 CNRS – Subscribe to ERCIM News by sending an email to Université Toulouse 3) [email protected] 12 Building Brains EditorialBoard: by Steve Furber (The University of Manchester) Central editor: Peter Kunz, ERCIM office ([email protected]) 14 The BrainScaleS Accelerated Analogue Neuromorphic Architecture Local Editors: by Johannes Schemmel (Heidelberg University, • Christine Azevedo Coste, Inria, France ([email protected]) Germany) • Andras Benczur, SZTAKI, Hungary ([email protected]) • José Borbinha, Univ. of Technology Lisboa, Portugal ([email protected]) 15 BrainScaleS: Greater Versatility for Neuromorphic • Are Magnus Bruaset, SIMULA, Norway ([email protected]) Emulation • Monica Divitini, NTNU, Norway ([email protected]) by Andreas Baumbach (Heidelberg University, • Marie-Claire Forgue, ERCIM/W3C ([email protected]) University of Bern), Sebastian Billaudelle (Heidelberg • Lida Harami, FORTH-ICT , Greece ([email protected]) University), Virginie Sabado (University of Bern) and • Athanasios Kalogeras, ISI, Greece ([email protected]) Mihai A. Petrovici (University of Bern, Heidelberg • Georgia Kapitsaki, Univ. of Cyprus, Cyprus ([email protected]) University) • Annette Kik, CWI, The Netherlands ([email protected]) • Hung Son Nguyen, Unviv. of Warsaw, Poland ([email protected]) 17 Fast and Energy-efficient Deep Neuromorphic • Alexander Nouak, Fraunhofer-Gesellschaft, Germany Learning ([email protected]) by Julian Göltz (Heidelberg University, University of • Maria Rudenschöld, RISE, Sweden ([email protected]) Bern), Laura Kriener (University of Bern), Virginie • Harry Rudin, Switzerland ([email protected]) Sabado (University of Bern) and Mihai A. Petrovici • Erwin Schoitsch, AIT, Austria ([email protected]) (University of Bern, Heidelberg University) • Thomas Tamisier,LIST, Luxembourg ([email protected]) • Maurice ter Beek, ISTI-CNR, Italy ([email protected]) 19 Higher Cognitive Functions in Bio-Inspired Artificial Intelligence Cover photo by Josh Riemer on Unsplash. by Frédéric Alexandre, Xavier Hinaut, Nicolas Rougier and Thierry Viéville (Inria)

2 ERCIM NEWS 125 April 2021 20 Self-Organizing Machine Architecture 39 NEUROTECH - A European Community of Experts by Bernard Girau (Université de Lorraine), Benoît on Neuromorphic Technologies Miramond (Université Côte d’Azur), Nicolas Rougier by Melika Payvand, Elisa Donati and Giacomo Indiveri (Inria Bordeaux) and Andres Upegui (University of (University of Zurich) Applied Sciences of Western Switzerland) 40 NeuroAgents – Autonomous Intelligent Agents that 22 Reentrant Self-Organizing Map: Toward Brain- Interact with the Environment in Real Time inspired Multimodal Association by Giacomo Indiveri (University of Zurich and ETH by Lyes Khacef (University of Groningen), Laurent Zurich) Rodriguez and Benoît Miramond (Université Côte d’Azur, CNRS) RESEARCH ANd INNOvATION 24 Brain-inspired Learning Drives Advances in Neuromorphic Computing 42 Human-like AI by Nasir Ahmad (Radboud University), Bodo by Dave Raggett, (W3C/ERCIM) Rueckauer (University of Zurich and ETH Zurich) and Marcel van Gerven (Radboud University) 44 Graph-based Management of Neuroscience Data: Representation, Integration and Analysis 25 Memory Failures Provide Clues for more Efficient by Maren Parnas Gulnes (University of Oslo / SINTEF Compression AS), Ahmet Soylu (OsloMet – Oslo Metropolitan by Dávid G. Nagy, Csenge Fráter and Gergő Orbán University) and Dumitru Roman (SINTEF AS) (Wigner Research Center for Physics) 45 The ICARUS Ontology: A General Aviation 27 Neuronal Communication Process Opens New Ontology Directions in Image and Video Compression Systems by Joanna Georgiou, Chrysovalantis Christodoulou, by Effrosyni Doutsi (FORTH-ICS), Marc Antonini George Pallis and Marios Dikaiakos (University of (I3S/CNRS) and Panagiotis Tsakalides (University of Cyprus) Crete and FORTH/ICS) 47 Security Management and the Slow Adoption of 28 Fulfilling Brain-inspired Hyperdimensional Blockchains Computing with In-memory Computing by Peter Kieseberg, Simon Tjoa and Herfried Geyer (St. by Abbas Rahimi, Manuel Le Gallo and Abu Sebastian Pölten University of Applied Sciences) (IBM Research Europe) 48 Trick the System: Towards Understanding 30 E = AI 2 Automatic Speech Recognition Systems by Marco Breiling, Bijoy Kundu (Fraunhofer Institute by Karla Markert (Fraunhofer AISEC) for Integrated Circuits IIS) and Marc Reichenbach (Friedrich-Alexander-Universität Erlangen-Nürnberg) 50 ACCORDION: Edge Computing for NextGen Applications 32 Brain-inspired Visual-Auditory Integration Yielding by Patrizio Dazzi ) Near Optimal Performance – Modelling and Neuromorphic Algorithms by Timo Oess and Heiko Neumann (Ulm University) ANNOuNCEMENTS

34 Touch in Robots: A Neuromorphic Approach 51 Dagstuhl Seminars and Perspectives Workshops by Ella Janotte (Italian Institute of Technology), Michele Mastella Elisabetta Chicca (University of 51 W3C Workshop on Wide Color Gamut and High Groningen) and Chiara Bartolozzi (Italian Institute of Dynamic Range for the Web Technology) 51 W3C Workshop on Smart Cities 35 Uncovering Neuronal Learning Principles through Artificial Evolution 51 Horizon Europe Project Management by Henrik D. Mettler (University of Bern), Virginie Sabado (University of Bern), Walter Senn (University of Bern), Mihai A. Petrovici (University of Bern and Heidelberg University) and Jakob Jordan (University of Bern)

37 What Neurons Do – and Don’t Do by Martin Nilsson (RISE Research Institutes of Sweden) and Henrik Jörntell (Lund University, Department of Experimental Medical Science)

ERCIM NEWS 125 April 2021 3 Joint ERCIM Actions

• Establishing the advanced disaster reduction management system by fusion of real-time disaster simulation and big data assimilation by Prof. Shunichi Koshimura, Interna - tional Research Institute of Disaster Science,Tohoku Univ. • Exploring etiologies, sub-classification, and risk predic - tion of diseases based on big-data analysis of clinical and ERCIM-JST Joint whole omics data in medicine, by Prof. Tatsuhiko Tsuno - da, Graduate School of Science, Univ. of Tokyo Symposium on Big data • Early detection and forecasting of pandemics using large- scale biological data: designing intervention strategy by and Artificial Intelligence Prof. Kimihito Ito (on Account for Hiroshi Nishiura (PI)), Division of Bioinformatics, Research Center for Zoonosis ERCIM and Japan Science and Technology Agency (JST) Control, Hokkaido Univ. have organised the first joint symposium on Big Data and • Statistical computational cosmology with big astronomi - Artificial Intelligence, held on 18 and 19 February 2021. The cal data by Prof. Naoki Yoshida, Department of online symposium aimed to present recent results of research Physics/Kavli IPMU, Univ. of Tokyo conducted in the context of the JST CREST programme, as • Data-driven analysis of the mechanism of animal develop - well as relevant research results from European institutions. ment by Shuichi Onami, RIKEN Center for Biosystems The meeting provided the opportunity to the participants Dynamics Research from Japan and Europe to familiarize themselves with recent • Knowledge Discovery by Constructing AgriBigData by research results and consider collaboration prospects that Masayuki Hirafuji, Graduate School of Agricultural and will arise in the context of the Horizon Europe framework Life Sciences, Univ. of Tokyo programme or relevant initiatives from JST. • Scientific Paper Analysis: Knowledge Discovery through Approximatively 80 invited participants attended the event. Structural Document Understanding by Yuji Matsumoto, RIKEN Center for Advanced Intelligence Project. Ms. Hiroko Tatesawa, Manager, ICT Group, Department of Strategic Basic Research, JST, introduced JST. JST, as one of The CREST programme is one of JST’s major undertakings the national Japanese funding agencies has about 1200 for stimulating achievement in fundamental science fields. employees and an annual budget of 122.4 billion Yen (about 950 million Euro), of which about two thirds are allocated to From ERCIM, three 20 minutes presentations were given: funding programmes including strategic basic research, • Scalable computing infrastructures: Keeping up with data international collaborations, industry-acedemia collabora - growth by Prof. Angelos Bilas, FORTH tion and technology transfer. ERCIM President Dr. Björn • Security, Privacy and Explainability for ML Models by Levin (RISE), then gave an overview on the activities of Prof. Andreas Rauber, SBA Research and TU Wien and by ERCIM. Dr. Rudolf Mayer, SBA Research • IoT - Big Data, Privacy: Resolving the contradiction by With nine short talks, JST delivered an overview of CREST Rigo Wenning, Legal counsel, W3C. Program on Big Data Applications: • Development of a knowledge-generating platform driven The first day concluded with a presentation “AI and HPC” by big data in drug discovery through production process - given by Prof. Satoshi Matsuoka, Director, RIKEN Center es by Dr. Makoto Taiji, (on Account for Kimito Funatsu for Computational Science. In his talk, Prof. Matsuoka pre - (PI)) sented “Fugaku”, the largest and fastest exascale supercom - • Big data assimilation and AI: Creating new development puter, how it is built and used for applications and experi - in real-time weather prediction by Dr. Takemasa Miyoshi, ments. These include for example the exploration of new RIKEN Center for Computational Science drug candidates for COVID-19, a digital twin of the entire smart city of Fugaku as a shareable smart city R&D environ -

Prof.YuzuruTanaka,HokkaidoUniversity(left)andProf.DimitrisPlexousakis,FORTH-ICSandUniversityofCrete.Screenshotstakenduring theseminar.

4 ERCIM NEWS 125 April 2021 ment, and unprecedented deep learning scalability for AI ERCIM “Alain Bensoussan” developments. Fellowship Programme The second day started with three more 20 minutes presenta - tions from European researchers: The ERCIM PhD Fellowship Programme has been • Big Data Trends and Machine Learning in Medicinal established as one of the premier activities of ERCIM. The Chemistry by Prof. Jürgen Bajorath, Prof., Univ. Bonn programme is open to young researchers from all over the • Data security and privacy in emerging scenarios by Prof. world. It focuses on a broad range of fields in Computer Sara Foresti, Univ. Milano Science and Applied Mathematics. • Using Narratives to Make Sense of Data by Dr. Carlo Meghini, CNR. The fellowship scheme also helps young scientists to improve their knowledge of European research structures Dr. Katsumi Emura, Director of AIP Network Laboratory, and networks and to gain more insight into the working con - JST gave at talk “Promoting innovation with advanced ditions of leading European research institutions. research ecosystem - Introduction of JST AIP Network Laboratory”. The AIP Network Laboratory is a virtual labo - The fellowships are of 12 months duration (with a possible ratory with teams in the fields of AI, big data, IoT, and cyber extension), spent in one of the ERCIM member institutes. security under the JST Strategic Basic Research Program. Fellows can apply for second year in a different institute. This talk was followed by five short overviews of some other JST CREST Programs in the ICT area: • Introduction for the research area of Core Technologies for Advanced Big Data Integration Program by Prof. Masaru ¶¶The fellowship provided me a great Kitsuregawa, Director-General, National Institute of platform to collaborate internationally Informatics, Univ. of Tokyo with the proficient researchers working • A Future Society Allowing Human-Machine Harmonious in my domain and sharpen my research skills. It has broadened my career Collaboration by Prof. Norihiro Hagita, Art Science options both in academia and industry Department, Osaka University of Arts and augmented the chances of getting • Symbiotic Interaction Research toward Society of Human my dream jobs. and AI by Prof. Kenji Mase Graduate School of Informat - ics, Nagoya Univ. Nancy AGARWAL Former ERCIM Fellow • Towards the Next Generation IoT Systems by Prof. Emer - itus Hideyuki Tokuda, President, National Institute of Information and Communications Technology (NICT) • Fundamental Information Technologies toward Innovative Conditions Social System Design by Prof. Sadao Kurohashi, Gradu - Candidates must: ate School of Informatics. Kyoto Univ. • have obtained a PhD degree during the last eight years (prior to the year of the application deadline) or be in the Prof. Masashi Sugiyama, Director, RIKEN Center for last year of the thesis work with an outstanding academic Advanced Intelligent Project, Univ. of Tokyo then gave a record; keynote talk entitled “Fundamental Machine Learning • be fluent in English; Research Opens Up New AI Society”. He presented the • have completed their PhD before starting the grant; RIKEN Center for Advanced Intelligence Project and • submit the following required documents: cv, list of publi - machine learning as the core research challenge of current cations, two scientific papers in English, contact details of AI. two referees. The last presentation of the symposium was given by Dr. Fosca Giannotti, CNR on “Explainable Machine Learning The fellows are appointed either by a stipend (an agreement for Trustworthy AI”. for a research training programme) or a working contract. The symposium was organised and chaired by Prof. emeri - The type of contract and the monthly allowance/salary tus Yuzuru Tanaka, Hokkaido University, Prof. emeritus depends on the hosting institute. Nicolas Spyratos, Université Paris - Saclay, and Prof. Dim - itris Plexousakis, FORTH-ICS and University of Crete. Application deadlines They closed the symposium with the wish to continue the Deadlines for applications are currently 30 April and 30 cooperation between ERCIM and JST. September each year.

Since its inception in 1991, over 500 fellows have passed More information: through the programme. In 2020, 22 young scientists com - Programme and links to slides and recorded presentations: menced an fellowship and 77 fellows have been hosted https://www.ercim.eu/events/ercim-jst-joint-symposium during the year. Since 2005, the programme is named in honour of Alain Bensoussan, former president of Inria, one Please contact: of the three ERCIM founding institutes. Prof. Dimitris Plexousakis, Director FORTH-ICS, Greece http://fellowship.ercim.eu

ERCIM NEWS NEWS 125 April 2021 5 Special Theme

Introduction to the Special Theme Brain-inspired Computing –

by Robert Haas (IBM Research Europe) and Michael Pfeiffer (Bosch Center for Artificial Intelligence)

The origins of artificial intelligence (AI) Samsung and Huawei have focused on novel fabric optimised to transmit can be traced back to the desire to build disruptive hardware technologies and spikes, SpiNNaker allowed the first thinking machines, or electronic brains. systems. In technology roadmaps, real-time simulation of a model of the A defining moment occurred back in brain-inspired computing is commonly early sensory cortex, estimated to be 1958, when Frank Rosenblatt created seen as a future key enabler for AI on two to three times faster than when the first artificial neuron that could learn the edge. using GPUs or HPC systems. by iteratively strengthening the weights BrainScaleS introduced novel chips of of the most relevant inputs and Artificial neural networks (ANNs) have analogue hardware circuits to represent decreasing others to achieve a desired only vague similarities with biological each neuron, which are then configured output. The IBM 704, a computer the neurons, which sparingly communicate in spiking neural networks, performing size of a room, was fed a series of punch using binary spikes at a low firing rate. over 1000 times faster than real time, cards and after 50 trials it learnt to dis - However, compared with conventional and leverage general-purpose cores tinguish cards marked on the left from ANNs, spiking neural networks (SNNs) closest to the analogue cores to control cards marked on the right. This was the require further innovation to be able to the plasticity of each neuron's synapses. demonstration of the single-layer per - classify data accurately. Wozniak et al. This enables experimentation with var - ceptron, or, according to its creator, "the (page 8) present the spiking neural unit ious learning methods: For instance, first machine capable of having an orig - (SNU), a model that can be introduced Baumbach et al. (page 15) describe use inal idea". [1] into deep learning architectures and cases that exploit these cores and simu - trained to a high accuracy over a broad late the environment of a bee looking Computation in brains and the creation range of applications, including music for food, or perform reinforcement of intelligent systems have been studied and weather prediction. Yin et al. (page learning for the Pong game. In Göltz et in a symbiotic fashion for many decades. 9) show another application of back - al. (page 17), the timing of spikes is This special theme highlights this propagation through time for tasks such exploited, with an encoding repre - enduring relationship, with contribu - ECG analysis and speech recognition, senting more prominent features with tions from some of the region’s leading with a novel type of adaptive spiking earlier spikes, and training of the academic and industrial research labora - recurrent neural network (SRNN) that synaptic plasticity implemented by tories. Ongoing collaboration between achieves matching accuracy and an esti - error backpropagation of first-time neuroscientists, cognitive scientists, AI mated 30 to 150-fold energy improve - spikes. researchers, and experts in disruptive ment over conventional RNNs. computing hardware technologies and Masquelier (page 11) also exploits How can modern AI benefit from neu - materials has made Europe a hotspot of backpropagation through time, but roscience and cognitive science brain-inspired computing research. The accounts for the timing of spikes too, research? Alexandre et al. (page 19) progress has been accelerated by EU- achieving comparable classification present their interdisciplinary approach funded activities in the Future and accuracy of speech commands com - towards transferring neuroscientific Emerging Technologies (FET) pro - pared with conventional deep neural findings to new models of AI. Two arti - gramme, and more recently in the networks. cles demonstrate how both hardware- Human Brain Project. Over the last and data-efficiency can be increased by decade, similar large-scale interdiscipli - SpiNNaker, described by Furber (page following brain-inspired self-organisa - nary projects have started in the rest of 12), and BrainScaleS, outlined by tion principles. The SOMA project, pre - the world, such as the BRAIN initiative Schemmel (page 14), are the two neuro - sented by Girau et al. (page 20), applies in the US, and national initiatives in morphic computing systems that are both structural and synaptic neural plas - China, Canada and Australia. Brain- being developed within the Human ticity principles to 3D cellular FPGA inspired computing projects from large Brain Project. With a million cores rep - platforms. The multi-modal Reentrant industrial players such as IBM, Intel, resenting neurons interconnected by a Self-Organizing Map (ReSOM) model

6 ERCIM NEWS 125 April 2021 presented by Khacef et al. (page 22) Janotte et al. (page 34) describe a neuro - inspired algorithms, sensors, and highlights new opportunities to reduce morphic approach to both sensing and processors. the need for high volumes of labelled local processing of tactile signals. In data by exploiting multi-modal associa - their approach, event-driven sensors in Although AI-based systems have tions. Ahmad et al. (page 24) introduce the electronic skin encode tactile reached performance levels beyond more plausible approaches towards stimuli, and spiking neural networks human capabilities in many specialised plasticity to replace the well-known, but learn to extract information that can be domains, such as image classification biologically unrealistic, backpropaga - used for real time, low power control of and game playing, it’s important to note tion algorithm used for training deep robots or prostheses. that there is still a long way to go for us neural networks. Nagy et al. (page 25) to reach the flexibility and efficiency of use results from human memory experi - Computational neuroscience is an biological brains in real-world domains. ments to inform a new semantic com - important source of information for Building increasingly brain-like AI also pression technique for images, which brain-inspired computing. When trying requires re-thinking fundamental prin - captures the gist of visual memories. to model the vast repertoire of learning ciples of computation: the classical von- The highly efficient processing of the rules at play at any time in different Neumann model with its separation of visual system is used as inspiration for a parts of the brain, manual modeling of computation and memory is not an effi - novel image and video compression plasticity rules beyond Hebbian cient way to implement large scale bio - technique by Doutsi et al. (page 27), learning and STDP becomes tedious. logically inspired neural networks. exploiting the temporal dynamics of Mettler et al. (page 35) demonstrate Recent progress in neuromorphic com - spiking neurons to detect characteristics how new mechanistic principles and puting platforms and in-memory com - of visual scenes. rules for plasticity can be discovered by puting provides new opportunities for evolutionary algorithms, using an fast, energy-efficient, massively par - A promising approach to brain-like evolving-to-learn approach. A new allel, and continuously learning large- computation, which could be applied to mechanistic model of the cerebellar scale brain-inspired computing systems machine learning and robotics, is com - Purkinje neuron, by Nilsson and based on spiking neural networks. puting with very high-dimensional, Jörntell (page 37), is matched to in-vivo Furthermore, envisioning more unreliable, highly efficient, and widely recordings of spike communications advanced capabilities, such as making distributed neuronal representations. and provides a surprisingly simple analogies and reasoning on problems is Rahimi et al. (page 28) present an explanation as to what such cells actu - another step that will hinge on an inter - implementation of hyperdimensional ally do. play of ANNs and symbolic AI, a recent computing on integrated memristive and very active field of research. crossbars, showing how this computa - Europe is currently a leader in tional paradigm is particularly well- advancing neuromorphic computing Reference: suited for memristive hardware. Other technologies, but there is still a long [1] Melanie Lefkowitz, “Professor´s examples of hardware–software co- road ahead to translate the research and perceptron paved the way for AI – design are described by Breiling et al. technology development results into 60 years too soon”, Cornell (page 30), who present the results of a novel products. The NEUROTECH Chronicle, https://kwz.me/h5j national competition in Germany, in project, presented by Payvand et al. which entrants developed energy-effi - (page 39), has been very successful in Please contact: cient AI hardware for ECG classifica - building a community for European Robert Haas tion. researchers in the field of neuromorphic IBM Research Europe, Zurich computing. Among the highlights are Research Laboratory, Switzerland Brains naturally combine signals from monthly educational events on core [email protected] different modalities, which can help themes such as event-driven sensing, develop better sensor fusion algorithms and circuits for processing and learning. Michael Pfeiffer for artificial autonomous systems. A Projects such as NeuroAgents, Bosch Center for Artificial neuromorphic implementation of a described by Indiveri (page 40), aim to Intelligence, Renningen, Germany visual-auditory integration model using develop autonomous intelligent agents [email protected] bio-inspired sensors for accurate and that show cognitive abilities on robotic stable localisation is presented by Oess platforms, entirely based on brain- and Neumann (page 32). Similarly,

ERCIM NEWS 125 April 2021 7 Special Theme

Biologically Inspired dynamics Enhance deep Learning

by Stanisław Woźniak, Thomas Bohnstingl and Angeliki Pantazi (IBM Research – Europe)

Deep learning has achieved outstanding success in several artificial intelligence (AI) tasks, resulting in human-like performance, albeit at a much higher power than the ~20 watts required by the human brain. We have developed an approach that incorporates biologically inspired neural dynamics into deep learning using a novel construct called spiking neural unit (SNU). Remarkably, these biological insights enabled SNU-based deep learning to even surpass the state-of-the-art performance while simultaneously enhancing the energy-efficiency of AI hardware implementations.

The state-of-the-art deep learning is widely considered as the most prom - ticular, we demonstrated that the leaky based on artificial neural networks ising contender for the next generation integrate and fire (LIF) evolution of the (ANNs) that only take inspiration from of neural networks [1]. However, mod - membrane potential of a biological biology to a very limited extent – prima - elling and training complex SNNs has neuron (Figure 1a) can be modelled, rily from its networked structure. This remained a challenge, which has led to maintaining the exact equivalence, has several drawbacks, especially in the conviction among many researchers through a specific combination of recur - terms of power consumption and energy that the performance of ANNs is in prac - rent artificial neurons forming jointly an efficiency [1]. Our group at IBM tice superior to that of SNNs. SNU (Figure 1b). Therefore, SNUs can Research – Europe has been studying be flexibly incorporated into deep more realistic models of neural dynamics We have developed a novel spiking learning architectures and trained to in the brain, known as spiking neural net - neural unit (SNU) that unifies the bio - high accuracy using supervised methods works (SNNs). Since human brains are logically inspired dynamics of spiking developed for ANNs, such as the back - still much more capable than modern neurons with the research advances of propagation through time algorithm. We deep learning systems, SNNs have been ANNs and deep learning [2, L1]. In par - have compared SNUs and common deep

Biology Deep Learning

Rate-coded MNIST classification SNU-based in-memory computing

0 1 0 0 1 1 ... 0 0 1 0

Feed-forward network architectures

. .

. . . .

IIn-memn-me oryr computationtaationn . . . . .

x . . 1 . . s . . e

g a x

t 2 . . l . .

o . .

.

v . . . PCM PCM ... PCM ...... t . . cells cells cells u p

n

I - +

.

. . xm 2-PCPCCM . - + In

- + SNU

I1 I2 In I 2 .

.

. ... SN SNU SNU SNU I1 U Training an arcchitecturehitecture of 784-256-256-10 neurons SNU 600 d

TTestest error (Hardware) o o h 8x h c i l e o e k p i l m - i e 400 t

g

l o g l l

] n a e i s v [ i w n t i

a a g U r e t P

200 n

r e C e g p a

1x r e v 0 A

Figure1:Incorporatingbiologicallyinspireddynamicsintodeeplearning:a.Biologicalneuronsreceiveinputspikesthataremodulatedby synapticweightsatthedendritesandaccumulatedintothemembranepotentialVmincellsoma.Thisistypicallymodelledasaresistor-capacitor (RC)circuit.Theoutputspikesareemittedthroughaxonstodownstreamneurons.b.SNUmodelsthespikingneuraldynamicsthroughtwo recurrentartificialneurons.N1performstheaccumulationandcorrespondstotheVm.N2controlsthespikeemissionandresetstheVm.c.Digit classificationforrate-codedMNISTdataset.Intheupperpartofthepane,feed-forwardnetworksofcommondeeplearningunitsarecompared. Higheraccuracyindicatesimprovement.Inthelowerpartofthepane,thetrainingtimeofSNUvs.LSTMisillustrated.d.Thesynaptic operationsareacceleratedin-memorythroughphysicalpropertiesofthecrossbarstructure,illustratedintheupperpartofthepane.Weusetwo PhaseChangeMemorydevicespersynapse.Thelowerpartofthepanecontainsacomparisonoftheaveragenegativelog-likelihoodofsoftware simulationvs.hardwareexperimentforthemusicpredictiontaskusingJSBdataset.Lowervaluescorrespondtohigher-qualitypredictions. Figureadaptedfrom[2].

8 ERCIM NEWS 125 April 2021 learning units, such as LSTMs and The increasing adoption of ANNs has unlock the potential of neuromorphic GRUs, on tasks including image classi - sparked the development of hardware hardware by training deep SNNs to high fication (Figure 1c, upper part), lan - accelerators to speed up the required accuracy. Our team is continuing the guage modelling, music prediction and computations. Because SNUs cast the work towards developing the biologi - weather prediction. Importantly, we dynamics of spiking neurons into the cally inspired deep learning paradigm demonstrated that although SNUs have deep learning framework, they provide by also exploring ways of enhancing the the lowest number of parameters, they a systematic methodology for training learning algorithms with neuroscientific can surpass the accuracy of these units SNNs using such AI accelerators. We insights [4]. and provide a significant increase in demonstrated this with a highly effi - speed (Figure 1c, lower part). Our cient in-memory computing concept Link: results established the SNN state of the based on nanoscale phase-change [L1] https://kwz.me/h5b art with the best accuracy of 99.53% +/- memory devices (Figure 1d), where the 0.03%, for the handwritten digit classifi - spiking nature of SNNs leads to further References: cation task based on the rate-coded energy savings [3]. Furthermore, we [1] W. Maass: “Networks of spiking MNIST dataset using a convolutional showed that SNNs are robust to opera - neurons: The third generation of neural network. Moreover, we even tion with low-precision synaptic neural network models,” Neural demonstrated the first-of-a-kind tem - weights down to 4-bits of precision and Networks, vol. 10, no. 9, pp. poral generative adversarial network can also cope with hardware imperfec - 1659–1671, 1997. (GAN) based on an SNN. tions such as noise and drift. [2] S. Woźniak, et al.: “Deep learning incorporating biologically inspired Neural circuits in the human brain Our work on SNU has bridged the ANN neural dynamics and in-memory exhibit an additional set of functionali - and the SNN worlds by incorporating computing,” Nat Mach Intell, vol. ties observed on the neural dynamics biologically inspired neural dynamics 2, no. 6, pp. 325–336, Jun. 2020. level, where adaptive spiking thresholds into deep learning, enabling to benefit [3] A. Pantazi, et al.: “All-memristive play an important role as well as on the from both worlds. SNU allows SNNs to neuromorphic computing with architectural level, where lateral inhibi - take direct advantage of recent deep level-tuned neurons,” tion between neurons is a commonly learning advances, which enable easy Nanotechnology, vol. 27, no. 35, p. observed theme. Another advantage of scaling up and training deep SNNs to a 355205, 2016. the SNUs is that the developed frame - high degree of accuracy. From the ANN [4] T. Bohnstingl, et al.:, “Online work can easily be extended to incorpo - perspective, SNU implements novel Spatio-Temporal Learning in Deep rate such neural functionalities. For temporal dynamics for machine Neural Networks,” example, we demonstrated that the LI- learning applications with fewer param - https://arxiv.org/abs/2007.12723 SNU variant, that implements the lat - eters and potentially higher perform - eral inhibition, achieves digit classifica - ance than the state-of-the-art units. Please contact: tion performance that is on par with the Finally, hardware accelerators can also Stanisław Woźniak, IBM Research – original SNU while significantly benefit from SNUs, as they allow for a Europe, Switzerland reducing the required number of spikes. highly efficient implementation and [email protected]

Effective and Efficient Spiking Recurrent Neural Networks by Bojian Yin (CWI), Federico Corradi (IMEC Holst Centre) and Sander Bohté (CWI)

Although inspired by biological brains, the neural networks that are the foundation of modern artificial intelligence (AI) use exponentially more energy than their counterparts in nature, and many local “edge” applications are energy constrained. New learning frameworks and more detailed, spiking neural models can be used to train high-performance spiking neural networks (SNNs) for significant and complex tasks, like speech recognition and EGC-analysis, where the spiking neurons communicate only sparingly. Theoretically, these networks outperform the energy efficiency of comparable classical artificial neural networks by two or three orders of magnitude.

Modern deep neural networks have only formed by synapses, and the means of neurons could be constructed that would vague similarities with the functioning of communication: real neurons communi - be at least as powerful as similar sized biological neurons in animals. As illus - cate not with analogue values, but with networks of standard analogue artificial trated in Figure 1 (left), many details of isomorphic pulses, or spikes, and they neurons [1]. Additionally, the sparse biological neurons are abstracted in this do so only sparingly. nature of spike-based communication is model of neural computation, such as the likely key to the energy-efficiency of spatial extent of real neurons, the vari - As far back as 1997, Maass argued that biological neuronal networks, an able nature and effect of real connections mathematically, networks of spiking increasingly desirable property as artifi -

ERCIM NEWS 125 April 2021 9 Special Theme

Figure1:Left:operationofspikingneuronscomparedtoanalogueneuralunits.Neuronscommunicatewithspikes.Eachinputspikeaddsa weighted(anddecaying)contributiontotheinternalstateofthetargetedneuron.Whenthisstateexceedssomethresholdfrombelow,aspikeis emitted,andtheinternalstateisreset.Intheanalogueneuralunit,inputsxareweightedwithcorrespondingweightswtoaddtotheneuron’s internalstate,fromwhichtheneuron’soutputG(x)iscomputed.Right:illustrationofaspikingrecurrentneuralnetwork.Redconnectionsdenote theeffectiveself-recurrenceofspikingneuronsduetothedynamicinternalstate.

cial intelligence (AI) energy consump - train these networks to solve hard We believe this work demonstrates that tion is rapidly escalating and many edge- benchmarks using Back-Propagation- SNNs can be compelling propositions AI applications are energy-constrained. Through-Time: the SRNN is illustrated for many edge-AI applications in the Until recently, however, the accuracy of in Figure 1 (right). In particular, we form of neuromorphic computing. At the spiking neural networks (SNNs) was showed how an adaptive version of the same time, tunable spiking neuron poor compared to the performance of classical “leaky-integrate-and-fire” parameters proved essential for modern deep neural networks. spiking neuron, the Adaptive LIF achieving such accuracy, suggesting (ALIF) neuron, enables the networks to both novel venues for improving SNNs A principal obstacle for designing effec - achieve high performance by co- further and neuroscientific investigations tive learning rules in SNNs has always training the internal parameters of the into corresponding biological plasticity. been the discontinuous nature of the spiking neuron model that determine its spiking mechanism: the workhorse of temporal dynamics (the membrane- Links: deep learning is error-backpropagation, potential time-constant and the adapta - Project: NWO TTW programme which requires the estimation of the tion time-constant) together with the “Efficient Deep Learning”, Project 7, local gradient between cause and effect. weights in the networks. https://efficientdeeplearning.nl/ For spiking neurons, this requires, for Pre-print: https://kwz.me/h5t example, estimating how changing an With these adaptive SRNNs, we Code: https://kwz.me/h5d input weight into a neuron affects the achieve new state-of-the-art perform - emission of spikes. This, however, is ance for SNNs on several tasks of References: discontinuous: the spiking mechanism is inherent temporal nature, like ECG [1] W. Maass: “Fast sigmoidal networks triggered when the input into the neuron analysis and speech recognition (SHD), via spiking neurons. Neural exceeds an internal threshold, and a all while demonstrating highly sparse Computation”, 9(2), 279-304, 1997. small change in weight may be enough communication. Their performance also [2] E.O.Neftci, H. Mostafa, F. Zenke: to just trigger a spike or prevent a spike compared favourably to standard recur - “Surrogate gradient learning in from being emitted. To overcome the rent neural networks like LSTMs and spiking neural networks: Bringing problem of learning with such discontin - approaches the current state of the art in the power of gradient-based uous gradients, recent work [2] demon - deep learning. When comparing com - optimization to spiking neural strated a generic approach by using a putational cost as a proxy for energy networks”, IEEE Signal Processing “surrogate gradient”. Importantly, the consumption, we calculated the Magazine, 36(6), 51-63, 2019. surrogate gradient approach enables the required Multiply-Accumulate (MAC) [3] B. Yin, F. Corradi, S.M. Bohté: use of modern deep learning software and Accumulate (AC) operations for “Effective and efficient frameworks like PyTorch and various networks. With a MAC being computation with multiple- Tensorflow, including their efficient about 31x more energetically expensive timescale spiking recurrent neural optimisers and GPU-support. compared to an AC, and with spiking networks”, in Int. Conf. on neurons using ACs sparingly where Neuromorphic Systems 2020 (pp. In joint work from CWI and IMEC [3], MACs are used in standard neural net - 1-8), 2020; Arxiv 2005.11633v1 we demonstrate how surrogate gradi - works, we demonstrate theoretical ents can be combined with efficient energy advantages for the SRNNs Please contact: spiking neuron models in recurrent ranging from 30 to 150 for comparable Sander Bohte, CWI, The Netherlands spiking neural networks (SRNNs) to accuracy (for more details, see [3]). [email protected]

10 ERCIM NEWS 125 April 2021 Back-propagation Now Works in Spiking Neural Networks! by Timothée Masquelier (UMR5549 CNRS – Université Toulouse 3)

Back-propagation is THE learning algorithm behind the deep learning revolution. Until recently, it was not possible to use it in spiking neural networks (SNN), due to non-differentiability issues. But these issues can now be circumvented, signalling a new era for SNNs.

Biological neurons use short electrical desired output of the network. The deep networks, which has led to the impulses called “spikes” to transmit error, i.e., the distance between the widely talked about deep learning revo - information. The spike times, in addition actual and desired outputs, can be com - lution. to the spike rates, are known to play an puted on these labelled examples. important role in how neurons process Gradient descent is used to find the This has motivated us and others to train information. Spiking neural networks parameters of the networks (e.g., the SNNs with BP. But unfortunately, it is (SNNs) are thus more biologically real - synaptic weights) that minimise this not straightforward. To compute the istic than the artificial neural networks error. The strength of BP is to be able to gradients, BP requires differentiable (ANNs) used in deep learning, and are compute the gradient of the error with activation functions, whereas spikes are arguably the most viable option if one respect to all the parameters in the inter - “all-or-none” events, which cause dis - wants to understand how the brain com - mediate “hidden” layers of the network, continuities. Here we present two recent putes at the neuronal description level. whereas the error is only measured in methods to circumvent this problem. But SNNs are also appealing for AI, the output layer. This is done using a especially for edge computing, since recurrent equation, which allows com - S4NN: a latency-based they are far less energy hungry than putation of the gradients in layer l-1 as a backpropagation for static stimuli ANNs. Yet until recently, training SNNs function of the gradients in layer l. The The first method, S4NN, deals with with back-propagation (BP) was not gradients in the output layer are static stimuli and rank-order-coding [1]. possible, and this has been a major straightforward to compute (since the With this sort of coding, neurons can impediment to the use of SNNs. error is measured there), and then the fire at most one spike: most activated computation goes backward, until all neurons first, while less activated neu - Back-propagation (BP) is the main gradients are known. BP thus solves the rons fire later, or not at all. In particular, supervised learning algorithm in ANNs. “credit assignment problem”, i.e., it in the readout layer, the first neuron to Supervised learning works with exam - finds the optimal thing to do for the fire determines the class of the stimulus. ples for which the ground truth, or hidden layers. Since the number of Each neuron has a single latency, and “label”, is known, which defines the layers is arbitrary, BP can work in very we demonstrated that the gradient of the loss with respect to this latency can be approximated, which allows estimation of the gradients of the loss with respect to all the weights, in a backward manner, akin to traditional BP. This approach reaches a good accuracy, although below the state-of-the-art: e.g., a test accuracy of 97.4% for the MNIST dataset. However, the neuron model we use, non-leaky integrate-and-fire, is simpler and more hardware friendly than the one used in all previous similar proposals.

Surrogate Gradient Learning: a general approach One of the main limitations of S4NN is the at-most-one-spike-per-neuron con - straint. This constraint is acceptable for static stimuli (e.g., images), but not for those that are dynamic (e.g., videos, sounds): changes need to be encoded by additional spikes. Can BP still be used in this context? Yes, if the “surrogate gradient learning” (SGL) approach is Figure1.(Top)Exampleofspectrogram(Melfilters)extractedfortheword“off”.(Bottom) used [2]. Correspondingspiketrainsforonechannelofthefirstlayer.

ERCIM NEWS 125 April 2021 11 Special Theme

The most commonly used spiking be done in space, time (which simulates Neural Networks with One Spike neuron model, the leaky integrate-and- conduction delays), or both. We vali - per Neuron,” Int. J. Neural Syst., fire neuron, obeys a differential equa - dated our approach on a speech classifi - vol. 30, no. 06, p. 2050027, Jun. tion, which can be approximated using cation benchmark: the Google speech 2020. discrete time steps, leading to a recur - commands dataset. We managed to [2] E. O. Neftci, H. Mostafa, and F. rent relation for the potential. This rela - reach nearly state-of-the-art accuracy Zenke, “Surrogate Gradient tion can be computed using the recur - (94.5%) while maintaining low firing Learning in Spiking Neural rent neural network (RNN) formalism, rates (about 5Hz, see Figure 1). Our Networks,” IEEE Signal Process. and the training can be done using back- study has just been accepted at the IEEE Mag., vol. 36, no. October, pp. propagation through time, the reference Spoken Language Technology 51–63, 2019. algorithm for training RNNs. The firing Workshop [4]. Our code is based on [3] S. Woźniak, A. Pantazi, T. threshold causes optimisation issues, PyTorch and is available in open source Bohnstingl, and E. Eleftheriou, but they can be overcome by using a at [L1]. “Deep learning incorporating “surrogate gradient”. In short, in the biologically inspired neural forward pass of BP, the firing threshold Conclusion dynamics and in-memory is applied normally, using the Heaviside We firmly believe that these results computing,” Nat. Mach. Intell., step function. But in the backward pass, open a new era for SNNs, in which they vol. 2, no. 6, pp. 325–336, 2020. we pretend that a sigmoid was used in will compete with conventional deep [4] T. Pellegrini, R. Zimmer, and T. the forward pass instead of the learning in terms of accuracy on chal - Masquelier, “Low-activity Heaviside function, and we use the lenging problems, while their imple - supervised convolutional spiking derivative of this sigmoid for the gra - mentation on neuromorphic chips could neural networks applied to speech dient computation. This approximation be much more efficient and use less commands recognition,” in IEEE works very well in practice. In addition, power. Spoken Language Technology the training can be done using auto - Workshop, 2021. matic-differentiation tools such as Link: PyTorch or Tensorflow, which is very [L1] https://kwz.me/h5c Please contact: convenient. Timothée Masquelier References: Centre de Recherche Cerveau et We extended previous approaches by [1] S. R. Kheradpisheh and T. Cognition, UMR5549 CNRS – adding convolutional layers (see [3] for Masquelier, “Temporal Université Toulouse 3, France a similar approach). Convolutions can Backpropagation for Spiking [email protected]

Building Brains

by Steve Furber (The University of Manchester)

SpiNNaker – a Spiking Neural Network Architecture – is the world’s largest neuromorphic computing system. The machine incorporates over a million ARM processors that are connected through a novel communications fabric to support large-scale models of brain regions that operate in biological real time, with the goal of contributing to the scientific Grand Challenge to understand the principles of operation of the brain as an information processing system.

SpiNNaker is one of two neuromorphic In many ways SpiNNaker resembles a in the brain, where each neuron typi - computing systems offering an open massively-parallel high-performance cally connects to many thousands of service under the auspices of the EU computer (HPC), with two exceptions: other neurons. Flagship Human Brain Project (HBP) as (i) the processors used on SpiNNaker part of the EBRAINS research infra - are small, efficient cores intended for User access to SpiNNaker is typically structure [L1] – the other being the embedded applications rather than the mediated through the PyNN (Python BrainScaleS machine developed by the high-end powerful numerical proces - Neural Networks) language, which is University of Heidelberg in Germany. sors used in HPC, and (ii) the communi - supported by a number of simulators, so The machine has been 20 years in con - cations fabric on SpiNNaker is opti - models can be developed on any com - ception and 15 years in construction at mised for carrying large numbers of puter or laptop and then run on the University of Manchester, UK, and small packets (where each packet con - SpiNNaker, either through a web-based has been supporting a service with a veys a neuron “spike”) to many destina - batch mode interface or using a Jupyter half-million core system since April tions, whereas HPC systems are opti - notebook. For real-time robot control it 2016, upgraded to the full million cores mised for large, fast data movements is generally necessary to have a since November 2018. Around 100 addi - between two cores. The SpiNNaker SpiNNaker system collocated with the tional small SpiNNaker systems have communications architecture is very robot, though the host server in been deployed in research labs around much motivated by the very high degree Manchester supports the HBP the world. of connectivity found between neurons

12 ERCIM NEWS 125 April 2021 Figure1:Themillion-core SpiNNakermachineat Manchester.Themachine occupies10rackcabinets, eachwith120SpiNNaker boards,powersupplies, coolingandnetwork switches.The11thcabinet ontherightcontains serversthatsupport configurationsoftware andremoteaccesstothe machine.

Neurorobotics platform for virtual neu - butions from many other disciplines. References: rorobotic control simulation. SpiNNaker offers a platform for testing [1] Oliver Rhodes et al, “Real-time theories and hypotheses at scale as they cortical simulation on Examples of brain models successfully emerge from such work, and also for neuromorphic hardware”, Phil. run on the machine include a cortical investigating how concepts from brain Trans. R. Soc. A. 378:20190160. microcircuit [1], where SpiNNaker was science can be translated into advances http://doi.org/10.1098/rsta.2019.01 the first machine to achieve real time, in artificial intelligence with applica - 60 and a hybrid model of the auditory tions in the commercial domain. [2 ] Robert James et al, “Parallel system [2]. Further models are being Distribution of an Inner Hair Cell developed in collaboration with part - The technology employed in the current and Auditory Nerve Model for ners within the HBP. SpiNNaker machine is now 10 years Real-Time Application”. IEEE old, and its successor has been under Trans. Biomed. Circuits Syst. Even with a million processor cores, development within the HBP as a col - 12(5):1018-1026. doi: SpiNNaker cannot model systems laboration between the University of 10.1109/TBCAS.2018.2847562 approaching the scale of the full human Manchester and the Technical [3] Steve Furber & Petrut Bogdan brain with its (just under) one hundred University of Dresden. SpiNNaker2 (eds.) (2020), “SpiNNaker: A billion neurons and several hundred tril - will be fabricated in 2021 and will Spiking Neural Network lion synapses, but it could potentially deliver an order of magnitude increase Architecture”, Boston-Delft: now support models of the scale of the in functional density and energy effi - publishers, mouse brain, which is about a thousand ciency, taking SpiNNaker forward into http://dx.doi.org/10.1561/97816808 times smaller than the human brain. its third decade, but with an increased 36523 However, biological data is relatively emphasis on commercial applications sparse, and putting together a full brain alongside brain science. Please contact: model even of an insect’s brain, which Steve Furber is a hundred to a thousand times smaller The first 20 years of the SpiNNaker The University of Manchester, UK than a mouse brain, involves interpreta - project have been documented in an [email protected] tion of, and extrapolation from, the Open Access book [3], including details available data, combined with more of a number of applications and the cur - than a little guesswork! So progress rent state of development of towards a reasonably coherent descrip - SpiNNaker2. tion of information processing in the brain will be a long, slow haul, Links: requiring diligent work from practical [L1] https://kwz.me/h5y and theoretical neuroscientists, [L2] https://kwz.me/h5H advances in brain imaging, and contri -

ERCIM NEWS 125 April 2021 13 Special Theme

The BrainScaleS Accelerated Analogue Neuromorphic Architecture

by Johannes Schemmel (Heidelberg University, Germany)

By incorporating recent results from neuroscience into analogue neuromorphic hardware we can replicate biological learning mechanisms in silicon to advance bio-inspired artificial intelligence.

Biology is still vastly ahead of our engi - nervous system. Acceleration factors Suitable tasks range from fundamental neering approaches when we look at from 1000 to 100,000 have been neuroscientific research, like the valida - computer systems that cope with natural achieved in previous implementations, tion of results from experimental obser - data. Machine learning and artificial thereby dramatically shortening execu - vations in living tissue, to AI applica - intelligence increasingly use neural net - tion times. tions like data classification and pattern work technologies inspired by nature, recognition. but most of these concepts are founded The BrainScaleS architecture consists on our understanding of biology from of three pillars: These custom-developed plasticity pro - the 1950s. The BrainScaleS neuromor - • microelectronics as the physical sub - cessing units have access to a multitude phic architecture builds upon contempo - strate for the emulation of neurons of analogue variables characterising the rary biology to incorporate the knowl - and synapses state of the network. For example, the edge of modern neuroscience within an • a software stack controlling the oper - temporal correlation of action potentials electronic emulation of neurons and ation of a BrainScaleS neuromorphic is continuously measured by the ana - synapses [1]. system from the user interface down logue circuitry and can be used as input to the individual transistors variables for learning algorithms [3]. Biological mechanisms are directly • training and learning algorithms sup - This combination of analogue and dig - transferred to analogue electronic cir - ported by dedicated software and ital hardware-based learning is called cuits, representing the signals in our hardware throughout the system. hybrid plasticity. brains as closely as possible with respect to the constraints of energy and area effi - The analogue realisation of the network Figure 1 shows the recently finished ciency. Continuous signals like voltages emulation does not only have the BrainScaleS mobile system. A and currents at cell membranes are advantage of speed, but also consumes BrainScaleS ASIC is visible on the directly represented as they evolve over much less energy and is easier and lower left side of the topmost circuit time [2]. The communication between therefore cheaper to fabricate than com - board. This version is suited for edge- neurons can be modelled by either faith - parable digital implementations. It does based AI applications. Edge-based AI fully reproducing each action potential not need the smallest possible transistor avoids the transfer of full sensor infor - or by summarising them as firing rates, size to achieve a high energy efficiency. mation to the data centre and thus as is commonly done in contemporary To configure the analogue circuits for enhances data privacy and has the artificial intelligence (AI). their assigned tasks, they are accompa - potential to vastly reduce the energy nied by specialised microprocessors consumption of the data centre and data Using the advantages of modern micro - optimised for the efficient implementa - communication infrastructure. An ini - electronics, the circuits are operated at a tion of modern bio-inspired learning tial demonstration successfully detects much higher speed than the natural algorithms. signs of heart disease in EKG traces.

Figure1:Themostrecentadditiontothe BrainScaleSfamilyofanalogueaccelerated neuromorphichardwaresystemsisacredit- cardsizedmobileplatformforEdgeAI applications.Theprotectivelidabovethe BrainScaleSASIChasbeenremovedforthe photograph.

14 ERCIM NEWS 125 April 2021 The energy used to perform this opera - References: [3] S. Billaudelle et al.: “Structural tion is so low that a wearable medical [1] S. A. Aamir et al.: “A Mixed- plasticity on an accelerated analog device could work for years on battery Signal Structured AdEx Neuron for neuromorphic hardware system,” power using the BrainScaleS ASIC. Accelerated Neuromorphic Cores,” Neural Netw., vol. 133, pp. 11–20, IEEE Trans. Biomed. Circuits Oct. 2020, doi: The BrainScaleS architecture can be Syst., vol. 12, no. 5, pp. 10.1016/j.neunet.2020.09.024. accessed and tested free of charge 1027–1037, Oct. 2018, doi: online at [L1]. 10.1109/TBCAS.2018.2848203. Please contact: [2] S. A. Aamir, Y. Stradmann, and P. Johannes Schemmel This work has received funding by the Müller: “An accelerated lif Heidelberg University, Germany, EC (H2020) under grant agreements neuronal network array for a large- [email protected] 785907 and 945539 (HBP) and the scale mixed-signal neuromorphic Björn Kindler BMBF (HD-BIO-AI, 16ES1127) architecture,” on Circuits and …, Heidelberg University, Germany, 2018, [Online]. Available: [email protected] Link: [L1] https://kwz.me/h5E https://kwz.me/h5K.

BrainScaleS: Greater versatility for Neuromorphic Emulation by Andreas Baumbach (Heidelberg University, University of Bern), Sebastian Billaudelle (Heidelberg University), Virginie Sabado (University of Bern) and Mihai A. Petrovici (University of Bern, Heidelberg University)

Uncovering the mechanics of neural computation in living organisms is increasingly shaping the development of brain-inspired algorithms and silicon circuits in the domain of artificial information processing. Researchers from the European Human Brain project employ the BrainScaleS spike-based neuromorphic platform to implement a variety of brain-inspired computational paradigms, from insect navigation to probabilistic generative models, demonstrating an unprecedented degree of versatility in mixed-signal neuromorphic substrates.

Unlike their machine learning counter - tions. These features, including the on- exploiting the speed of the accelerated parts, biological neurons interact prima - chip learning capabilities of the second platform for the experiment. In total, the rily via electrical action potentials, generation of the BrainScaleS architec - emulated insect performs about 5.5 bio - known as “spikes”. The second genera - ture, have the potential to enable the logical hours’ worth of exploration in tion of the BrainScaleS neuromorphic widespread use of spiking neurons 20 seconds. system [1, L6] implements up to 512 beyond the realm of neuroscience and such spiking neurons, which can be into artificial intelligence (AI) applica - While accelerated model execution is near-arbitrarily connected. In contrast to tions. Here we showcase the system’s certainly an attractive feature, it is often classical simulations, where the simula - capabilities with a collection of five the learning process that is prohibitively tion time increases for larger systems, highly diverse emulation scenarios, time-consuming. Using its on-chip gen - this form of in-silico implementation from a small model of insect navigation, eral-purpose processor, the BrainScaleS replicates the physics of neurons rather including the sensory apparatus and the system offers the option of fully than numerically solving the associated environment, to fully fledged discrimi - embedded learning, thus maximising equations, enabling what could be natory and generative models of higher the benefit of its accelerated neuro- described as perfect scaling, with the brain functions [1]. synaptic dynamics. Using the dedicated duration of an emulation being essen - circuitry for spike-timing-dependent tially independent of the size of the As a first use case, we present a model of plasticity (STDP), the system can, for model. Moreover, the electronic circuits the bee’s brain that reproduces the ability example, implement a neuromorphic are configured to be approximately 1000 of bees to return to their nest’s location agent playing a version of the game times faster than their biological coun - after exploring their environment for Pong. The general-purpose processor terparts. It is this acceleration that makes food (Figure 1). The on-chip general- again simulates the game dynamics and, BrainScaleS-2 a powerful platform for purpose processor of BrainScaleS is used depending on the game outcome, pro - biological research and potential appli - to simulate the environment and to stim - vides the success signal used in the rein - cations. Importantly, the chip was also ulate the motor neurons triggering the forcement learning. Together with the designed for energy efficiency, with a exploration of the environment. The net - analogue STDP measurements, this total nominal power consumption of work activity stores the insect’s position, reward signal allows the emulated agent 1 W. Coupled with its emulation speed, enabling autonomous navigation back to to learn autonomously. The resulting this means energy consumption is sev - its nest. The on-chip environment simu - system is an order of magnitude faster eral orders of magnitude lower than lation avoids delays typically introduced and three orders of magnitude more state-of-the art conventional simula - by the use of coprocessors, thus fully

ERCIM NEWS 125 April 2021 15 Special Theme

Figure1:(A)PhotographoftheBrainScaleSchipwithfalse-colouroverlayofsomeofitscomponents.(B)100pathstravelledbyanemulated insectswarmingoutfromitsnest,exploringtheworld(grey)andreturninghome(pink).(C)EmulatedPongagentandassociatedspiking network.Emulationtimeandenergyconsumptioncomparedtostate-of-theartsimulationsshownin(D)and(E).

energy-efficient than an equivalent soft - two-dimensional dataspace (Figure 2). of weaker synaptic connections. On ware simulation (Figure 1). For each neuron, the number of possible BrainScaleS, it is possible to route the inputs is limited by the architecture and output spikes of multiple neurons to In a different example of embedded topology of the silicon substrate, much each synaptic circuit. This allows the learning, a network is trained to dis - like the constraints imposed by biolog - adaptation of the network connectome criminate between different species of ical tissue. Here too, the learning rule is at runtime. In particular, synapses iris flowers based on the shape of their based on spike-time similarities but below a threshold strength are periodi - petals using a series of receptors in a enhanced by regularisation and pruning cally removed and alternatives are instantiated. The label neurons on BrainScaleS then develop distinct receptive fields using the same circuitry. For this particular problem, it was pos - sible to enforce a sparsity of nearly 90% without impacting the final classifica - tion performance (Figure 2).

The last two networks demonstrated in [1] hint at the more challenging nature of tasks that biological agents must solve in order to survive. One such task is being able to cope with incomplete or inaccurate sensory input. The Bayesian brain hypothesis posits that cortical activity instantiates probabilistic infer - Figure2:(A)DifferentspeciesofIriscanbedistinguishedbytheshapeoftheirflowers’petals. ence at multiple levels, thus providing a (B)Structuralchangestotheconnectomeofanemulatedclassifierorganizeappropriate normative framework for how mam - receptivefieldsevenforverysparseconnectivitythroughouttheexperiment.(C)Evolutionof malian brains can perform this feat. As classificationaccuracyduringtrainingfordifferentsparsitylevels. an exemplary implementation thereof, we showcase Bayesian inference in spiking networks on BrainScaleS (Figure 3). Similar systems have previ - ously been used to generate and dis - criminate between hand-written digits or small-scale images of fashion articles [3, 4].

The final network model was trained as a classifier for visual data using time-to- first-spike coding. This specific repre - sentation of information enabled it to classify up to 10,000 images per second using only 27 µJ per image. This partic - ular network is described in more detail in its own dedicated article [L5]. Figure3:Bayesianinferencewithspikingneuralnetworks.(A,B)Neuronalspikepatternscan beinterpretedasbinaryvectors.(C)Neuraldynamicscanlearntosamplefromarbitrary These five networks embody a diverse probabilitydistributionsoverbinaryspaces(D). set of computational paradigms and

16 ERCIM NEWS 125 April 2021 demonstrate the capabilities of the Meier, as well as the Manfred Stärk IEEE, 2020. BrainScaleS architecture. Its dedicated foundation for ongoing support. This [2] T. Wunderlich, et al.: spiking circuitry coupled with its versa - research has received funding from the “Demonstrating advantages of tile on-chip processing unit allows the European Union’s Horizon 2020 neuromorphic computation: a pilot emulation of a large variety of models in Framework Programme for Research study,” Frontiers in a particularly tight power and time enve - and Innovation under the Specific Grant Neuroscience,vol. 13, p. 260, 2019. lope. As work continues on user-friendly Agreement No. 945539 (Human Brain [3] D. Dold, et al. “Stochasticity from access to its many features, BrainScaleS Project SGA3). function—why the bayesian brain aims to support a wide range of users may need no noise.” Neural and use cases, from both the neuroscien - Links: networks 119 (2019): 200-213. tific research community and the [L1] https://youtu.be/_x3wqLFS278 [4] A. Kungl, F. Akos, et al.: domain of industrial applications. [L2] https://kwz.me/h5S “Accelerated physical emulation of [L3] https://kwz.me/h54 Bayesian inference in spiking We would like to thank our collaborators [L4] https://kwz.me/h50 neural networks”, Frontiers in Yannik Stradmann, Korbinian Schreiber, [L5] https://kwz.me/h51 neuroscience 13 (2019): 1201. Benjamin Cramer, Dominik Dold, Julian [L6] https://kwz.me/h52 Göltz, Akos F. Kungl, Timo C. Please contact: Wunderlich, Andreas Hartel, Eric References: Andreas Baumbach, NeuroTMA group, Müller, Oliver Breitwieser, Christian [1] S. Billaudelle, et al.: “Versatile Department of Physiology, University Mauch, Mitja Kleider, Andreas Grübl, emulation of spiking neural of Bern, Switzerland and Kirchhoff- David Stöckel, Christian Pehle, Arthur networks on an accelerated Institute for Physics, Heidelberg Heimbrecht, Philipp Spilger, Gerd neuromorphic substrate.” 2020 University, Germany Kiene, Vitali Karasenko, Walter Senn, IEEE International Symposium on [email protected] Johannes Schemmel and Karlheinz Circuits and Systems (ISCAS).

Fast and Energy-efficient deep Neuromorphic Learning by Julian Göltz (Heidelberg University, University of Bern), Laura Kriener (University of Bern), Virginie Sabado (University of Bern) and Mihai A. Petrovici (University of Bern, Heidelberg University)

Many neuromorphic platforms promise fast and energy-efficient emulation of spiking neural networks, but unlike artificial neural networks, spiking networks have lacked a powerful universal training algorithm for more challenging machine learning applications. Such a training scheme has recently been proposed and using it together with a biologically inspired form of information coding shows state-of-the-art results in terms of classification accuracy, speed and energy consumption.

Spikes are the fundamental unit in which University and computational neurosci - implements a version of error backprop - information is processed in mammalian entists at the University of Bern, fos - agation on first-spike times, which we brains, and a significant part of the infor - tered by the European Human Brain discuss in more detail below. mation is encoded in the relative timing Project. of these spikes. In contrast, the computa - This algorithm was demonstrated on the tional units of typical machine learning In the implementation on BrainScaleS- BrainScaleS-2 neuromorphic platform models output a numeric value without 2, to further enforce fast computation using both an artificial dataset resem - an accompanying time. This observation and to minimise resource requirements, bling the Yin-Yang symbol [3], as well is, in a modified form, at the centre of a an encoding was chosen where more as the real-world MNIST data set of new approach: a network model and prominent features are represented by handwritten digits. The Yin-Yang learning algorithm that can efficiently earlier spikes, as seen in nature, e.g., in dataset highlights the universality of the solve pattern recognition problems by how nerves in the fingertips encode algorithm and its interplay with first- making full use of the timing of spikes information about touch (Figure 1). spike coding in a small network, [1]. This quintessential reliance on From the perspective of an animal ensuring that training of this classifica - spike-based communication perfectly looking to survive, this choice of coding tion problem achieves highly accurate synergises with efficient neuromorphic is particularly appealing, as actions results (Figure 2). For the digit-recogni - spiking-network emulators, such as the must often be taken under time pres - tion problem, an optimised implementa - BrainScaleS-2 platform [2], thus being sure. The biological imperative of short tion yields particularly compelling able to fully harness their speed and times-to-solution is similarly applicable results: up to 10,000 images can be clas - energy characteristics. This work is the to silicon, carrying with it an optimised sified in less than a second at a runtime result of a collaboration between neuro - usage of resources. In this model of power consumption of only 270 mW, morphic engineers at the Heidelberg neural computation, synaptic plasticity which translates to only 27 µJ per

ERCIM NEWS 125 April 2021 17 Special Theme

system, or hardware dedicated to classi - fication alone will further exploit the algorithm’s benefits. Even though the current BrainScaleS-2 generation is limited in size, the algorithm can scale up to larger systems. In particular, cou - pled with the intrinsically parallel nature of the accelerated hardware, a scaled-up version of our model would not require longer execution time, thus conserving its advantages when applied to larger, more complex data. We thus view our results as a successful proof- Figure1:(Left)Discriminatornetworkconsistingofneurons(squares,circles,andtriangles) of-concept implementation, high - groupedinlayers.Informationispassedfromthebottomtothetop,e.g.pixelbrightnessofan lighting the advantages of sparse, but image.Here,adarkerpixelisrepresentedbyanearlierspike.(Right)Eachneuronspikesno robust coding combined with fast, low- morethanonce,andthetimeatwhichitspikesencodestheinformation. power silicon substrates, with intriguing potential for edge computing and neuro - prosthetics. image. For comparison, the power our approach is stable with regards to drawn by the BrainScaleS-2 chip for various forms of noise and deviations We would like to thank our collabora - this application is about the same as a from the ideal model, which represents tors Andreas Baumbach, Sebastian few LEDs. an essential prerequisite for physical Billaudelle, Oliver Breitwieser, computation, be it biological or artifi - Benjamin Cramer, Dominik Dold, Akos The underlying learning algorithm is cial. This makes our algorithm suitable F. Kungl, Walter Senn, Johannes built on a rigorous derivation of the for implementation on a wide range of Schemmel and Karlheinz Meier, as well spike time in biologically inspired neu - neuromorphic platforms. as the Manfred Stärk foundation for ronal systems. This makes it possible to ongoing support. This research has precisely quantify the effect of input Although these results are already received funding from the European spike times and connection strengths on highly competitive compared to other Union’s Horizon 2020 Framework later spikes, which in turn allows this related neuromorphic realisations of Programme for Research and effect to be computed throughout net - spike-based classification (Figure 3), it Innovation under the Specific Grant works of multiple layers (Figure 1). The is important to emphasise that the Agreement No. 945539 (Human Brain precise value of a single spike’s effect is BrainScaleS-2 neuromorphic chip is not Project SGA3). used on a host computer to calculate a specifically optimised for our form of change in the connectivity of neurons neural computation and learning but is Links: on the chip that improves the network’s rather a multi-purpose research device. [L1] https://kwz.me/h53 output. Crucially, we demonstrated that It is likely that optimisation of the [L2] https://kwz.me/h5S [L3] https://kwz.me/h54

References: [1] J. Göltz, L. Kriener, et al.: “Fast and deep: energy-efficient neuromorphic learning with first- spike times,” arXiv:1912.11443, 2019. [2] S. Billaudelle, et al.: “Versatile emulation of spiking neural Figure2:ArtificialdatasetresemblingtheYinYangsymbol[3],andtheoutputofanetwork networks on an accelerated trainedwithouralgorithm.Thegoalofeachofthetargetneuronsistospikeasearlyas neuromorphic substrate,” 2020 possible(smalldelay,brightyellowcolor)whenadatapoint,representedasacircle,is“intheir IEEE International Symposium on area”.Onecanseethatthethreeneuronscoverthecorrectareasinbrightyellow. Circuits and Systems (ISCAS). IEEE, 2020. [3] L. Kriener, et al.: “The Yin-Yang dataset,” arXiv:2102.08211, 2021.

Please contact: Julian Göltz, NeuroTMA group, Department of Physiology, University of Bern, Switzerland and Kirchhoff- Figure3:Comparisonwithotherspike-basedneuromorphicclassifiersontheMNISTdataset, Institute for Physics, Heidelberg see[1]fordetails.[A]:E.Stromatiasetal.2015,[B]:S.Esseretal.2015,[C]:G.Chenetal. University, Germany 2018. [email protected]

18 ERCIM NEWS 125 April 2021 Higher Cognitive Functions in Bio-Inspired Artificial Intelligence by Frédéric Alexandre, Xavier Hinaut, Nicolas Rougier and Thierry Viéville (Inria)

Major algorithms from artificial intelligence (AI) lack higher cognitive functions such as problem solving and reasoning. By studying how these functions operate in the brain, we can develop a biologically informed cognitive computing; transferring our knowledge about architectural and learning principles in the brain to AI.

Digital techniques in artificial intelli - to what is generally considered as between two kinds of loops. Loops gence (AI) have been making enormous human intelligence. The questions are involving agranular areas of the frontal progress and offer impressive perform - thus: how is it possible to implement cortex, which have been well studied in ance for the cognitive functions they prospective reasoning and planning rodents, are responsible for learning model. Deep learning has been primarily with these techniques? What about sensorimotor skills and stimulus-driven developed for pattern matching, and problem solving and creativity? decision-making. These loops can be extensions like Long Short Term Together with neuroscientists, we are related to elementary cognitive func - Memory (LSTM) networks can identify studying the principles of brain organi - tions described above, providing imme - and predict temporal sequences. sation and memory network interac - diate responses to external cues, and Adaptations to other domains, such as tions (Figure 1). The main goal of our internal reinforcement. In contrast, deep reinforcement learning, allow com - work, which takes place in the granular frontal regions present in pri - plex strategies of decision-making to be Mnemosyne Inria lab at the Bordeaux mates, which have not been so well learnt to optimise cumulated rewards. Neurocampus, is to transfer major studied, are generally associated with experimental findings to new models of what is called meta-cognition. Here, the Contrary to human performance, these AI, capable of such higher cognitive same learning principles are applied to techniques require large training corpora functions. the history of performance of the ele - and long training times. These chal - mentary functions to decide which of lenges can be partly addressed by We started by modelling the loops these elementary functions are trig - increasing data storage and computing between the frontal cortex and the basal gered, inhibited or updated, in a given power capacities. But another major ganglia, known to play a major role in context. This process yields contextual flaw is less often mentioned: all the decision-making and in the acquisition flexibility and rapid updating, as underlying cognitive functions are of skills [1]. This has led us to propose observed in higher cognitive functions. rather elementary and do not correspond that a major distinction should be made

Hippocampus CA3 Enthorinal Cortex

CA1 / Subiculum Dentate Gyrus Postrhinal Cortex Perirhinal Cortex

FrontalFrontal Cortex Sensory Cortex

ProprioceptionProprioception Location InteroceptionInteroception ExteroceptionExteroception vmPFC dmPFC dlPFC FEF motor Parietal Parietal Multimodal Insula TemporalTemporal association Like Need vlPFC Choose Do How WherWheree Why What

Motor Thalamus Sensory Thalamus

Basal Ganglia Extracortical

Shell CoreCore a m m l s s r s DMS DLS DLS s u u u u a l l -

NAcc NAcc o u u l l l l i r o d u u e e m m e p c c g a a i i b b l l l l y p y l l e VTAVTTAA SNC e a a u H o o r h h m S t t e C C A

VP GPi SNr C

Figure1:Thisglobalarchitectureofthebraincanbeconsideredtobemadeoffivecortico-basalloopsandtheirmainafferentsfrom hippocampusandotherextra-corticalstructures.Thethreemostcentralloops,whichimplementhighercognitivefunctions,includegranular frontalareas.

ERCIM NEWS 125 April 2021 19 Special Theme

From a modelling point of view, this semantic and procedural memory. In cognitive functions also corresponds to organisation of architecturally similar neuroscience, this has been shown to be manipulating structured knowledge rep - loops is very interesting because it a major way to accelerate learning for resentations and mechanisms of implies that similar computing princi - the latter kinds of memory. We are cur - abstract reasoning, which are classically ples (implemented in the circuitry of rently working on computational studied in symbolic AI. This opens cortico-basal loops) are exploited on models in this area, which are adapted unique perspectives into the study of the different kinds of information to imple - to both improve learning time and emergence of meaning and internal ment different kinds of behaviour decrease the size of the training corpora thought, which apply to the domains of (reflexive and reflective). We are, for needed. In addition, the hippocampus is natural language processing and educa - example, investigating the possibility of known to perform rapid and arbitrary tional science. basing decisions on the mechanisms of binding of information. It has been pro - working memory [2], which represents posed that the hippocampus in associa - Link: a history of past activity, instead of tion with the frontal cortex can build [L1] http://www.imn- decisions being made based on the prospective memory, also called imagi - bordeaux.org/en/teams/mnemosyne- activity itself. Likewise, decisions nation, which is fundamental for plan - mnemonic-synergy/ between different strategies can be ning in the future. This is another area made from levels of confidence esti - that we intend to study. References: mated from previous experiences, [1] T. Boraud, A. Leblois, N.P. instead of simply deciding from actual Here we have sketched a brief overview Rougier: “A natural history of rewards. Still relying on architectural of the cerebral architecture responsible skills”, Progress in Neurobiology, similarities, future work could be dedi - for higher cognitive functions and have 171, 114 124, 2018. cated to the processing of temporal explained that knowledge in neuro - https://doi.org/10.1016/j.pneurobio. sequences: Projections of agranular science is already available to transfer 2018.08.003 frontal areas to the parietal cortex computing principles to AI. [2] A. Strock, X. Hinaut, N.P. Rougier: implement the temporal chaining of Interestingly, this cerebral architecture “A Robust Model of Gated sensorimotor skills and homologous is tightly associated with brain regions Working Memory”, Neural projections from the granular frontal responsible for elementary cognitive Computation, 32(1), 153 181, 2019 areas are reported to play a role in functions related to sensorimotor pro - https://doi.org/10.1162/neco_a_012 sequential subgoal decomposition. cessing and reinforcement learning. 49 This research program can conse - [3] F. Alexandre: “A Global In addition, studying the projections quently be seen as an extension of framework for a systemic view of from the hippocampus to the cortico- existing models in AI. A more general brain modeling”, Brain Inf. 8, 3 basal loops has given us important reference to the nervous system [3] pro - (2021). insights that can help us address another vides the anchoring of these models in https://doi.org/10.1186/s40708- major problem related to prohibitive the perceptual, bodily, emotional and 021-00126-4 computational times for learning skills motivational dimensions of behaviour, and decision criteria. The hippocampus thus providing direct links for the Please contact: is reported to be the place of episodic robotic implementation of autonomous Frederic Alexandre learning, storing past episodes and agents, which we are also investigating. Inria Bordeaux Sud-Ouest, France learning to replay them to train On a different note, accessing higher [email protected]

Self-Organizing Machine Architecture

by Bernard Girau (Université de Lorraine), Benoît Miramond (Université Côte d’Azur), Nicolas Rougier (Inria Bordeaux) and Andres Upegui (University of Applied Sciences of Western Switzerland)

SOMA is a collaborative project involving researchers in France and Switzerland, which aims to develop a computing machine with self-organising properties inspired by the functioning of the brain. The SOMA project addresses this challenge by lying at the intersection of four main research fields, namely adaptive reconfigurable computing, cellular computing, computational neuroscience, and neuromorphic engineering. In the framework of SOMA, we designed the SCALP platform, a 3D array of FPGAs and processors permitting to prototype and evaluate self-organisation mechanisms on physical cellular machines.

The tremendous increase in transistor the available hardware resources. This drones. In the end, we cannot foresee integration in recent years has reached limit is even more acute when we con - every possible context that the system the limits of classic Von Neuman archi - sider application domains where the will face during its lifetime, thus tectures. Nonetheless, one major issue is system evolves under unknown and making it impossible to determine the the design and deployment of applica - uncertain conditions, such as mobile optimal hardware substrate. tions that cannot make optimal use of robotics, IoT, autonomous vehicles or Interestingly, the biological brain has

20 ERCIM NEWS 125 April 2021 solved this problem using a dedicated enough in the face of future massively architecture and mechanisms that offer parallel devices. We anticipate that the both adaptive and dynamic computa - expected properties of such alternative tions, namely, self-organisation [2]. Reconfigurable Cellular computing devices could emerge from a Computing Computing close interaction between cellular com - However, even though neuro-biological puting (decentralisation and hardware systems have often been a source of compliant massive parallelism) and inspiration for computer science, the neural computation (self-organisation transcription of self-organisation at the and adaptation). We also believe that hardware level is not straightforward, SOMA neuro-cellular algorithmics and hard - presenting several challenges. We are ware design are so tightly related that working on coupling this new computa - these two aspects should be studied tional paradigm with an underlying con - together. Therefore, we propose to com - ventional systolic architecture [1]. We Neuro Computational bine neural adaptivity and cellular com - Hardware Neuroscience use self-organised neural populations puting efficiency through a neuro-cel - on a cellular machine where local lular approach of synaptic and structural routing resources are not separated from self-organisation that defines a fully computational resources; this ensures Figure1:SOMAisaconvergencepoint decentralised control layer for neuro - natural scalability and adaptability as betweenpastresearchapproachestoward morphic reconfigurable hardware. To well as a better performance/power con - newcomputationparadigms:adaptive achieve this goal, the project brings sumption trade-off compared to other reconfigurablearchitecture,cellular together neuroscientists, computer sci - conventional embedded solutions. This computing,computationalneuroscience,and ence researchers, hardware architects new computing framework may indeed neuromorphichardware. and micro-electronics designers to represent a viable integration of neuro - explore the concepts of a Self- morphic computing into the classical Organizing Machine Architecture: Von Neumann architecture and could tures: the ability to focus on relevant SOMA (see Figure 1). This self-organi - endow these hardware systems with information, representation of informa - sation property is already being studied novel adaptive properties [3]. tion in a sparse manner, distributed data in various fields of computer science, processing and organisation fitting the but we are investigating it in an entirely Cortical plasticity and cellular nature of data, leading to a better effi - new context, applying perspectives computing in hardware ciency and robustness. Our goal is to from computational neuroscience to the This objective led us to study a desir - understand and design the first artificial design of reconfigurable microelec - able property from the brain that blocks that are involved in these princi - tronic circuits. The project focuses on encompasses all others: cortical plas - ples of plasticity. Hence, transposing the blocks that will pave the way in the ticity. This term refers to one of the plasticity and its underlying blocks into long term for smart computing sub - main developmental properties of the hardware is a step towards to defining a strates, exceeding the limits of current brain where the organisation of its struc - substrate of computation endowed with technology. ture (structural plasticity) and the self-organisation properties that stem learning of the environment (synaptic from the learning of incoming data. Convergence of research fields plasticity) develop simultaneously Previous work has explored the possi - toward an optimal computing effi - The neural principles of plasticity may bility of using neural self-organising ciency. This developmental process is not be sufficient to ensure that such a models to control task allocation on par - only made possible by some key fea - substrate of computation is scalable allel substrates [1], while adapting neural computational paradigms to cel - lular constraints. Adaptive reconfig - urable computing focuses on virtualisa - tion of reconfigurable hardware, run - time resource management, dynamic partial reconfiguration, and self-adap - tive architectures. Cellular approaches of distributed computing result in decentralised models that are particu - larly well adapted to hardware imple - mentations. However, cellular computa - tion still lacks adaptation and learning properties. This gap may be filled with the help of computational neuroscience and neuromorphic engineering through the definition of models that exhibit properties like unsupervised learning, Figure2:TheSCALPplatform,asetofFPGAsandprocessorswith3Dtopology,wasdesigned self-adaptation, self-organisation, and toevaluateself-organisationmechanismsoncellularmachines.Algorithmsbasedoncellular fault tolerance, which are of particular self-organisingmapsarethebasisoftheself-adaptationproperties. interest for efficient computing in

ERCIM NEWS 125 April 2021 21 Special Theme

embedded and autonomous systems. platform for implementing cellular neu - References: However, these properties only emerge romorphic architectures by imposing [1] L. Rodriguez, B. Miramond, B. from large fully connected neural maps real physical connectivity constraints. Granado “Toward a sparse self- that result in intensive synaptic commu - Also, a real 3D machine architecture organizing map for neuromorphic nications. (instead of a simulated one) can better architectures”, ACM JETC, 11 (4), handle scalability issues when model - 1-25, 2015. Our self-organising models are deployed ling dynamic bio-inspired 3D neural [2] G.I. Detorakis, N.P. Rougier: “A on the Self-configurable 3-D Cellular connectivity. We have already proposed Neural Field Model of the multi-FPGA Adaptive Platform such models using both dynamical Somatosensory Cortex: Formation, (SCALP) (Figure 2), which has been sprouting and pruning of synapses inside Maintenance and Reorganization of developed in the framework of the a self-organising map and a method to Ordered Topographic Maps”, PLoS SOMA project. SCALP is a multi-FPGA migrate neurons between clusters to ONE 7, 7, e40257, 2012. hardware platform that enables the cre - dynamically reassign neurons from one [3] A. Upegui, et al.: “The Perplexus ation of prototype 3D NoC architectures task to another. These methods provide bio-inspired reconfigurable circuit”, with dynamic topologies. A node is dynamic structural and computational in Proc. of the Second NASA/ESA composed of a Xilinx Zynq SoC (dual- resource allocations, inspired by the Conference on Adaptive Hardware core ARM Cortex-A9 @866 MHz + brain’s structural and functional plas - and Systems, 2007. Artix-7 programmable logic with 74,000 ticity, and are currently being deployed cells), 2 Gb DDR3 SDRAM, a 5-port onto the SCALP platform. Please contact: Ethernet switch, and a PLL. The Nicolas Rougier, Inria Bordeaux Sud- inherent cellular scalable architecture of Links: Ouest, France, SCALP, coupled with its enhanced inter - [L1] https://kwz.me/h55 [email protected] faces, provides the ideal computation [L2] https://kwz.me/h5 +33 7 82 50 31 10

Reentrant Self-Organizing Map: Toward Brain- Inspired Multimodal Association

by Lyes Khacef (University of Groningen), Laurent Rodriguez and Benoît Miramond (Université Côte d’Azur, CNRS)

Local plasticity mechanisms enable our brains to self-organize, both in structure and function, in order to adapt to the environment. This unique property is the inspiration for this study: we propose a brain-inspired computational model for self-organization, then discuss its impact on the classification accuracy and the energy-efficiency of an unsupervised multimodal association task.

Our brain-inspired computing approach emerging from local and dynamic inter - of the sensory modalities, such as sight, attempts to simultaneously reconsider actions, i.e., unifying structure and sound and touch, the brain arrives at AI and von Neumann's architecture. function in a single process: the plas - similar representations and concepts Both are formidable tools responsible ticity mechanism. It is hence the key to (convergence). On the other hand, bio - for digital and societal revolutions, but our ability to build our representation of logical observations show that one also intellectual bottlenecks linked to the the environment based on our experi - modality can activate the internal repre - ever-present desire to ensure the system ences, so that we may adapt to it. It is sentation of another modality. For is under control. The brain remains our also the basis of an extremely inter - example, when watching a specific lip only reference in terms of intelligence: esting characteristic of the human brain: movement without any sound, the we are still learning about its func - multimodal association. activity pattern induced in the early tioning, but it seems to be built on a very visual cortices activates in early audi - different paradigm in which its develop - In fact, most processes and phenomena tory cortices the representation of the mental autonomy gives it an efficiency in the natural environment are sound that usually accompanies the lip that we haven’t yet attained in com - expressed under different physical movement (divergence). puting. guises, which we refer to as different modalities. Multimodality is considered Here we summarise our work on the Our research focuses on the cortical a fundamental principle for the develop - Reentrant Self-Organizing Map plasticity that is the fundamental mecha - ment of embodied intelligence, as (ReSOM) [2], a brain-inspired compu - nism enabling the self-organization of pointed out by the neuroscientist A. tational neural system based on the the brain, which in turn leads to the Damasio, who proposed the reentry theory from G. Edelman [3] and emergence of consistent representations Convergence-Divergence Zone frame - J.P. Changeux using Kohonen Self- of the world. According to the neurobi - work [1]. Such a framework models the Organizing Maps (SOMs) and Hebbian- ologist F. Varela, self-organization can neural mechanisms of memorisation like learning to perform multimodal be defined as a global behaviour and recollection. Despite the diversity association (see Figure 1).

22 ERCIM NEWS 125 April 2021 Figure1:ReentrantSelf- OrganizingMap:(left) Processingpipelinefrom dataacquisitionatinputto multimodalassociationfor decisionmakingatthe outputwithunimodaland multimodalaccuraciesfora handgesturesrecognition taskbasedonaDVScamera andEMGsensor;(right) FPGA-basedneuromorphic implementationofthe proposedself-organizing artificialneuralnetworkon multipleSCALPboardsfor real-timeandenergy- efficientprocessing.

ReSOM model first labelled map. This way, the system between different modalities so that The brain’s plasticity can be divided breaks with the general principle of they complete each other and improve into two distinct forms: (i) structural classical machine learning by multimodal classification. Furthermore, plasticity, which, according to the exploiting the strength of the multi - it induces a form of hardware plasticity Selection Neural Groups Theory [3], modal association and takes advantage where the system’s topology is not fixed changes the neurons’ connectivity by of the coherence of the data from its by the user but learned along the sprouting (creating) or pruning experience to build in an incremental system’s experience through self-organ - (deleting) synaptic connections, and (ii) way a robust representation of the envi - ization. It reduces the inter-map com - synaptic plasticity that modifies ronment. From an application point of munication and thus reduces the (increases or decreases) the existing view, this means that the system only system’s energy consumption. This synaptic strength. We explore both needs few annotations from a single result could open up a whole lot of new mechanisms for multimodal association modality to label the maps of all the directions, inspired by the brain’s plas - through Hebbian-like learning. In the other modalities. Finally, once the mul - ticity, for future designs and implemen - resulting network, the excitement of timodal learning is done and all neurons tations of self-organizing hardware one part spreads to all the others and a from all SOMs are labelled, the system architectures in autonomous systems fragment of memory is enough to computes the convergence of the infor - such as vehicles, robots, drones or even awaken the entire memorised experi - mation from all the modalities to cortical prosthesis. ence. The network becomes both a achieve a better representation of the detector and a producer of signals. multi-sensory input. This global behav - Links: iour emerges from local interactions [L1] https://zenodo.org/record/4452953 First, the unimodal learning is per - among connected neurons. [L2] https://zenodo.org/record/3663616 formed independently for each modality using the SOM, a brain- Results and discussion References: inspired artificial neural network that Our experiments [2] show that the [1] Damasio: “Time-locked learns in an unsupervised manner divergence labelling leads to approxi - multiregional retroactivation: A (without labels). Then, based on co- mately the same unimodal accuracy as systems level proposal for the occurrent multimodal inputs, the neu - when using labels, while the conver - neural substrates of recall and rons of different SOMs create and rein - gence mechanism leads to a gain in the recognition”. force the reentrant multimodal associa - multimodal accuracy of +8.03% for a [2] Khacef et al.: “Brain-Inspired Self- tion via sprouting and Hebbian-like written/spoken digits database [L1] and Organization with Cellular learning. At the end of the multimodal +5.67% for a DVS/EMG hand gestures Neuromorphic Computing for binding, the neural group selection is database [L2]. We also gained +5.75% Multimodal Unsupervised made, and each neuron prunes up to when associating visual hand gestures Learning”. 90% of the possible connections to keep with spoken digits, illustrating the [3] Edelman: “Group selection and only the strongest ones. The third step is McGurk effect. Indeed, studies in cog - phasic reentrant signaling a theory then to give sense to these self-associ - nitive and developmental psychology of higher brain function”. ating groups of neurons. This is made show that spoken labels and auditory by labelling one of the SOMs maps modality in general add complementary Please contact: using very few labels (typically 1%), so information that improves object cate - Lyes Khacef, University of Groningen, that each neuron is assigned the class it gorisation. Netherlands, [email protected] represents. The fourth step is to label Prof. Benoît Miramond, Université the entire network (the other maps) by In summary, the ReSOM model Côte d’Azur, France, using the divergent activity from the exploits the natural complementarity [email protected]

ERCIM NEWS 125 April 2021 23 Special Theme

Brain-inspired Learning drives Advances in Neuromorphic Computing

by Nasir Ahmad (Radboud University), Bodo Rueckauer (University of Zurich and ETH Zurich) and Marcel van Gerven (Radboud University)

The success of deep learning is founded on learning rules with biologically implausible properties, entailing high memory and energy costs. At the Donders Institute in Nijmegen, NL, we have developed GAIT-Prop, a learning method for large-scale neural networks that alleviates some of the biologically unrealistic attributes of conventional deep learning. By localising weight updates in space and time, our method reduces computational complexity and illustrates how powerful learning rules can be implemented within the constraints on connectivity and communication present in the brain.

The exponential scaling of modern com - number of approaches have emerged tion of memory and compute, in con - pute power (Moore’s law) and data with candidate solutions to these prob - trast to traditional von-Neumann archi - storage capacity (Kryder’s law), as well lems. These include highly efficient tectures that are used by modern com - as the collection and curation of big hardware specifically designed to carry puters. Other key features include asyn - datasets, have been key drivers of the out the core tensor operations which chronous communication of the sub- recent deep learning (DL) revolution in compose ANN models (especially for processors (there is no global controller artificial intelligence (AI). This revolu - mobile devices), cloud-based compute of the system), and data-driven compu - tion, which makes use of artificial neural farms to supply on-demand compute to tation (computing only takes place with network (ANN) models that are trained internet-connected systems (held back significant changes in the input). A on dedicated GPU clusters, has afforded by access and relevant privacy con - number of companies and academic important breakthroughs in science and cerns) and more. research groups are actively pursuing technology, ranging from protein the development of such neuromorphic folding prediction to vehicle automa - Brain-inspired methods have also processors (Intel’s Loihi, IBM’s tion. At the same time, several out - emerged within this sphere. After all, TrueNorth, SpiNNaker, and standing challenges prohibit the use of the mammalian brain is a prime BrainScaleS, to name a few) [1]. These DL technology in resource-bounded example of a low-power and highly developments progress apace. applications. flexible information processing system. Neuromorphic computing is the name We expect that brain-inspired learning Among these issues, the quest for low- of the field dedicated to instantiating rules can become a major driver of latency, low-power devices with on- brain-inspired computational architec - future innovation for on-board neuro - demand computation and adaptability tures within devices. In general, neuro - morphic learning. In particular, the has become a field of competition. A morphic processors feature the co-loca - above described architectural design

Figure1:ComparisonofstandardbackpropagationwithourproposedGAIT-Propmethod.InsectionsA)andB),circlesindicatelayersofadeep

neuralnetwork.A)Inbackpropagation,allneuronactivationsy i ineachlayeriofthenetworkneedtobestoredduringaforwardpass,andthen thenetworkactivityhaltedsothatthatweightscanbeupdatedinaseparatebackwardpassbasedonaglobalerrorsignal.B)InGAIT-Prop,a top-downperturbationcircuitisdescribedwhichcantransmittargetvaluestirequiredtocomputeweightupdateslocallyateachunitbymaking useofthedynamicsofthesystem.C)Underparticularconstraints,GAIT-PropproducesidenticalweightupdatescomparedtoBackprop.D)This isalsoexhibitedduringtrainingwhereonarangeofbenchmarkdatasets(MNIST,KMNIST,andFashionMNIST)GAIT-Propmatchesthe performanceofbackpropagation.

24 ERCIM NEWS 125 April 2021 choices for neuromorphic chips pre - ular, our group recently developed a morphic systems as it ensures simplicity cisely match the constraints faced by method (GAIT-Prop [3]) to describe of node dynamics while enabling high neurons in brains (local memory and learning in ANN models. GAIT-Prop accuracy. compute, asynchronous communica - relies on the same system dynamics tion, data-driven computation and during inference and training (Figure We see our approach and extensions more). 1B) such that no additional machinery thereof, in which systems that learn are is required for gradient-based learning. close to indistinguishable in their Unfortunately, the computations When the system is provided with an dynamics to the systems carrying out required to carry out the traditional gra - indication of the “desired” output of the inference computations, as an important dient-based training of ANNs (known system (a training signal), it makes use step in the development of future neuro - as backpropagation of error) break the of theoretically determined inter-neuron morphic systems, as it mitigates the properties of both neuromorphic archi - connectivity to propagate this desired complexities associated with standard tectures and real neural circuits. Error signal across the network structure learning algorithms. Systems equipped computations in the typical format through the activities of the network with this capability could be embedded require non-local information, which units. The change in activity can then be in mobile devices and would be capable implies that the memory distributed used by each individual neuron to carry of learning with data locally, also across the sub-processing nodes would out relevant updates. thereby reducing privacy concerns need to communicate in a global which are common in an era of cloud- fashion (Figure 1A). For this reason Importantly, under some limited con - computing and mass data storage. alone, the traditional methods for back - straints, the updates produced by the propagation of error are undesirable for GAIT-Prop algorithm precisely match References neuromorphic “on-chip” training. those of the very powerful backpropa - [1] M. Bouvier et al.: “Spiking neural Furthermore, computations associated gation of error method (Figure 1C). networks hardware with learning and inference (i.e., the This ensures that we can achieve implementations and challenges: A application of the ANN) are carried out matching performance despite the local survey”, ACM JETC, 2019. in separate phases, leading to an unde - and distributed nature of the GAIT-Prop [2] T.P. Lillicrap et al.: sirable “blocking” phenomenon. By algorithm (Figure 1D). Our algorithm “Backpropagation and the brain”, comparison, the brain does not appear also provided for understanding how a Nat Rev Neurosci, 2020. to require non-local computations for desired network output can be trans - [3] N. Ahmad et al.: “GAIT-prop: A learning. Thus, by finding solutions to lated into target outputs for every node biologically plausible learning rule brain-inspired learning, we might arrive of an ANN system. Since our method derived from backpropagation of at solutions to “on-chip” training of relies on error signals being carried error”, NeurIPS, 2020. neuromorphic computing. within the dynamics of individual units of the network (requiring no specific Please contact: Recent developments in brain-inspired “error” nodes) it requires less computa - Nasir Ahmad learning have produced methods which tional machinery to accomplish Radboud University, Nijmegen meet these requirements [2]. In partic - learning. This feature is ideal for neuro - [email protected]

Memory Failures Provide Clues for more Efficient Compression by Dávid G. Nagy, Csenge Fráter and Gergő Orbán (Wigner Research Center for Physics)

Efficient compression algorithms for visual data lose information for curbing storage capacity requirements. An implicit optimisation goal for constructing a successful compression algorithm is to keep compression artifacts unnoticed, i.e., reconstructions should appear to the human eye to be identical to the original data. Understanding what aspects of stimulus statistics human perception and memory are sensitive to can be illuminating for the next generation of compression algorithms. New machine learning technologies promise fresh insights into how to chart the sensitivity of memory to misleading distortions and consequently lay down the principles for efficient data compression.

Humans are prone to errors when it memory has adapted to (natural statis - tional autoencoders, a new method in comes to recollecting details about past tics) so there is little data available machine learning that learns a latent experiences. Much research has about how these systematic failures link variable generative model of the data addressed the questions of which details to natural stimulus structure. statistics in an unsupervised manner [1]. our memory chooses to store and which Latent variable generative models aim are systematically discarded. Until Researchers at the Wigner Research to identify the features that contribute to recently we have not had methods to Center for Physics, Budapest, have the generation of the data and every learn the complex statistics to which addressed this challenge using varia - single data point is encoded as a combi -

ERCIM NEWS 125 April 2021 25 Special Theme

Figure1:Thedynamicsofforgetting:memoryundergoeschangesandinformationissystematicallydiscardedovertime,whichprovides insightsintothesensitivitiesofmemory.Anaccountofinformationlosswouldbetosimplylosememorythrough“fading”:uniformlylosing precisionwhenreconstructingepisodesfrommemory.Insteadofsuchnon-specificinformationloss,compressionwithlatent-variablegenerative modelsimpliesthatreconstructionerrorsreflectuncertaintyinlatentfeatures.Asthedelaybetweenencodingandrecollectionincreases,latent variablerepresentationofthestimulusisreshapedandwecancapturesignaturesoflower-ratecompressions.

nation of these features. The proposed same training objective as the theory References: semantic compression idea establishes that establishes the optimality criterion [1] D.G. Nagy, B. Török, G. Orbán: how to distinguish compression arte - for lossy compression. In fact, a con - “Optimal forgetting: Semantic facts that go unnoticed from those that tinuum of optimal compressions can be compression of episodic are more easily detected: artefacts that established, depending on the allowed memories”, PLoS Comp Biol, induce changes in the latent features are storage capacity, thus different genera - e1008367, 2020. more easily noticed than artefacts of tive models can be constructed for dif - [2] I. Higgins I, et al.: “beta-VAE: equal strength that leave the latent fea - ferent needs. Second, research by Nagy Learning Basic Visual Concepts tures unchanged. The theory of et al. provides support that the structure with a Constrained Variational semantic compression was put to a test of errors that generative compression is Framework”, Proc. of ICLR, 2018. by revisiting results from human sensitive to is similar to the structure of [3] A. Alemi, et al.: “An information- memory experiments in a wide array of errors that memory is sensitive to. theoretic analysis of deep latent- domains (recollection of words, chess Theory claims that this is not simply a variable models”, board configurations, or drawings) and lucky coincidence: the basis of it is that arXiv:1711:00464. showing that semantic compression can the generative model is trained on stim - predict how memory fails under a ulus statistics that are designed to Please contact: variety of conditions. approximate the statistics that the Gergő Orbán, Computational Systems human brain was adapted to. Neuroscience Lab, Dept. of Using generative models to encode Computational Sciences, Wigner complex stimuli offers new perspec - In the context of human memory, the Research Center for Physics, Hungary tives for compression. The power of study highlights a fascinating view. We [email protected] these generative models lies in the fact are all too aware that memories undergo that the latent features discovered by the disappointing dynamics that result in generative model provide a concise what we call “forgetting”. Is the gradual description of the data and are inher - process of forgetting best understood as ently empowered by reconstruction a simple process of fading? The answer capabilities [2]: in contrast to more tra - the authors provide could be more ditional deep learning models, varia - appropriately described as memories tional autoencoders are optimised for become less elaborate as time passes: efficient reconstruction ability. the theory of lossy compression estab - Nonlinear encoding is thus a way to lishes a harsh trade-off between storage build ultra-low bit-rate data compres - allowance and the richness of the sion technologies. This “generative retained details. As time passes it is compression” idea is boosted by two assumed that the capacity allocated for critical factors that concern two qualita - a memory decreases and details are lost tively different aspects of data compres - with stronger compression. And what sion. First, recent work in the field has are the details that remain? Those are demonstrated a tight link between varia - again determined by stimulus statistics: tional autoencoders and the theory of features that are closer to the gist of the lossy compression [3]. This link demon - memory prevail while specific details strates that VAEs are optimised for the disappear.

26 ERCIM NEWS 125 April 2021 Neuronal Communication Process Opens New directions in Image and video Compression Systems by Effrosyni Doutsi (FORTH-ICS), Marc Antonini (I3S/CNRS) and Panagiotis Tsakalides (University of Crete and FORTH/ICS)

The 3D ultra-high-resolution world that is captured by the visual system is sensed, processed and transferred through a dense network of tiny cells, called neurons. An understanding of neuronal communication has the potential to open new horizons for the development of ground-breaking image and video compression systems. A recently proposed neuro-inspired compression system promises to change the framework of the current state- of-the-art compression algorithms.

Over the last decade, the technological models that approximate neural behav - pression architectures. We developed a development of cameras and multimedia iour have been widely used for image neuro-inspired compression mechanism devices has increased dramatically to processing applications, including com - by using the Leaky Integrate-an-Fire meet societal needs. The significant pression. The biggest challenge how - (LIF) model, which is considered to be progress of these technologies has ush - ever, is that the brain uses the neural the simplest model that approximates ered in the big data era, accompanied by code to learn, analyse and make deci - neuronal activity, in order to transform serious challenges, including the sions without reconstructing the input an image into a code of spikes [1]. The tremendous increase in volume and visual stimulus. great advantage of the LIF model is that variety of measurements. Although most the code of spikes is generated as time big data challenges are being addressed There are several proven benefits to evolves, in a dynamic manner. An intu - with paradigm shifts in machine applying neuroscience models to com - itive explanation for the origin of this learning (ML) technologies, where a limited set of observations and associ - ated annotations are utilised for training models to automatically extract knowl - edge from raw data, little has been done about the disruptive upgrade of storage efficiency and compression capacity of existing algorithms.

The BRIEFING project [L1] aims to mimic the intelligence of the brain in terms of compression. The research is inspired by the great capacity of the visual system to process and encode Figure1:Anillustrationoftheneuro-inspiredcompressionmechanismthatenablesefficient visual information in an energy-efficient reductionofthenumberofbitsrequiredtostoreaninputimageusingtheLeakyIntegrate-and- and very compact yet informative code, Fire(LIF)modelasanapproximationoftheneuronalspikingactivity.Accordingtothis which is propagated to the visual cortex ground-breakingarchitecture,aninputimagecanbetransformedintoasequenceofspikes where the final decisions are made. whichareutilisedtostoreand/ortransmitthesignal.Theinterpretationofthespikesequence basedonsignalprocessingtechniquesleadstohighreconstructionqualityresults. If one considers the visual system as a mechanism that processes the visual stimulus, it seems an intelligent and very efficient model to mimic. Indeed, the visual system consumes low power, it deals with high resolution dynamic sig - nals (109 bits per second) and it trans - forms and encodes the visual stimulus in a dynamic way far beyond the current compression standards. During recent decades, there has been significant research into understanding how the visual system works, the structure and the role of each layer and individual cell that lies along the visual pathway, and Figure2:ThisgraphshowsthattheBRISQUEalgorithmthathasbeentrainedtodetectthe how the huge volume of visual informa - naturalcharacteristicsofvisualscenesisabletodetectfarmoreofthesecharacteristicswithin tion is propagated and compacted imagesthathavebeencompressedusingtheneuro-inspiredcompressionthantheJPEG through the nerve cells before reaching standards.TheBRISQUEscoresaretypicallybetween0and100,wherethelowerthescorethe the visual cortex. Some very interesting betterthenaturalcharacteristicsofthevisualscene.

ERCIM NEWS 125 April 2021 27 Special Theme

performance is that the longer the visual We have proven that neuro-inspired Systems Sophia Antipolis (I3S) at stimulus exists in front of a viewer, the compression is much more valuable French National Centre for Scientific better it is perceived. Similarly, the than the state-of-the-art such as JPEG Research (CNRS) longer the LIF model is allowed to pro - and/or JPEG2000, which both cause duce spikes, the more robust is the code. drastic changes in these features [2]. Link: This behaviour is far beyond the state- More specifically, we evaluated the [L1] https://kwz.me/h57 of-the-art image and video compression aforementioned models using the architectures that process and encode BRISQUE algorithm [3], a convolu - References: the visual stimulus immediately and tional neural network that has been pre- [1] E. Doutsi, et al.: “Neuro-inspired simultaneously without considering any trained in order to recognise natural Quantization”, 2018 25th IEEE time parameters. However, taking characteristics within the visual scene. International Conference on Image advantage of the time is very important, As a first step, we compressed a group Processing (ICIP), Athens, 2018, especially when considering a video of images with the same compression pp. 689-693. stream that is a sequence of images each ratio using both the neuro-inspired [2] E. Doutsi, L. Fillatre and M. of which exists for a given time. mechanism and the JPEG standard. Antonini: “Efficiency of the bio- Then, we fed the CNN with the com - inspired Leaky Integrate-and-Fire Another interesting aspect is that a pressed images and we observed that it neuron for signal coding”, 2019 neuro-inspired compression mechanism was able to detect far more natural char - 27th European Signal Processing can preserve important features in order acteristics within images that had been Conference (EUSIPCO), A Coruna, to characterise the content of the visual compressed by the neuro-inspired Spain, 2019, pp. 1-5. scene. These features are necessary for mechanism than JPEG standard. [3] A. Mittal, A. K. Moorthy, and A. several image analysis tasks, such as C. Bovik: “No-reference image object detection and/or classification. This project, funded by the French quality assessment in the spatial When the memory capacity or the band - Government within the framework of domain”, IEEE Transactions on width of the communication channel are the “Make Our Planet Green Again” Image Processing, vol. 21, no. 12, limited it is very important to transmit call, is a collaboration between the pp. 4695–4708, Dec 2012. the most meaningful information. In Institute of Compute Science (ICS) at other words, one needs to find the best Foundation for Research and Please contact: trade-off between the compression ratio Technology – Hellas (FORTH) and the Effrosyni Doutsi, ICS-FORTH, Greece and the image quality (rate-distortion). Laboratory of Informatics, Signals and [email protected]

Fulfilling Brain-inspired Hyperdimensional Computing with In-memory Computing

by Abbas Rahimi, Manuel Le Gallo and Abu Sebastian (IBM Research Europe)

Hyperdimensional computing (HDC) takes inspiration from the size of the brain’s circuits, to compute with points of a hyperdimensional space that thrives on randomness and mediocre components. We have developed a complete in-memory HDC system in which all operations are implemented on noisy memristive crossbar arrays while exhibiting extreme robustness and energy-efficiency for various classification tasks such as language recognition, news classification, and hand gesture recognition.

A cursory examination of the human The difference between traditional by the mathematical properties of brain shows: (i) the neural circuits are computing and HDC is apparent in the hyperdimensional spaces comprising very large (there can be tens of thou - elements that the machine computes high-dimensional binary vectors sands of fan-ins and fan-outs); (ii) with. In traditional computing, the ele - known as hypervectors [1]. activity is widely distributed within a ments are Booleans, numbers, and Hypervectors are defined as d- circuit and among different circuits; (iii) memory pointers. In HDC they are dimensional (where d ≥ 1,000) individual neurons need not be highly multicomponent vectors, or tuples, (pseudo)random vectors with inde - reliable; and (iv) brains operate with where neither an individual component pendent and identically distributed very little energy. These characteristics nor a subset thereof has a specific (i.i.d.) components. When the dimen - are in total contrast to the way tradi - meaning: a single component of a sionality is in the thousands, a huge tional computers are built and operate. vector and the entire vector represent number of quasi-orthogonal hypervec - Therefore, to approach such intelligent, the same thing. Furthermore, the vec - tors exist. This allows HDC to combine robust, and energy-efficient biological tors are very wide: the number of com - such hypervectors into new hypervec - computing systems, we need to rethink ponents is in the thousands. These tors using well-defined vector space and focus on alternative models of com - properties are based on the observation operations, defined such that the puting, such as hyperdimensional com - that key aspects of human memory, per - resulting hypervector is unique, and puting (HDC) [1][2]. ception, and cognition can be explained with the same dimension.

28 ERCIM NEWS 125 April 2021 Figure1:Theconceptofin-memoryHDC.Aschematicoftheconceptofin-memoryHDCshowingtheessentialstepsassociatedwithHDC(left) andhowtheyarerealizedusingin-memorycomputing(right).Anitemmemory(IM)storesh,d-dimensionalbasishypervectorsthatcorrespondto thesymbolsassociatedwithaclassificationproblem.Duringlearning,basedonalabelledtrainingdataset,adesignedencoderperforms dimensionality-preservingmathematicalmanipulationsonthebasishypervectorstoproducec,d-dimensionalprototypehypervectorsthatare storedinanAM.Duringclassification,thesameencodergeneratesaqueryhypervectorbasedonatestexample.Subsequently,anAMsearchis performedbetweenthequeryhypervectorandthehypervectorsstoredintheAMtodeterminetheclasstowhichthetestexamplebelongs.Inin- memoryHDC,boththeIMandAMaremappedontocrossbararraysofmemristivedevices.Themathematicaloperationsassociatedwith encodingandAMsearchareperformedinplacebyexploitingin-memoryreadlogicanddot-productoperations,respectively.Adimensionalityofd =10,000isused.SA,senseamplifier;ADconverters,analog-to-digitalconvertersareadaptedfrom[6].

HDC has been employed in a range of to many repetitive iterations in the deep tions implies that failure in a compo - applications, including traditional com - learning models [4]. nent of a hypervector is not contagious puting, machine learning, cognitive leading to robust computational frame - computing, and robotics [3]. It has HDC begins by representing symbols work. For instance, when unrelated shown significant promise in machine with i.i.d. hypervectors that are com - objects are represented by quasi- learning applications that involve tem - bined by nearly i.i.d.-preserving opera - orthogonal 10,000-bit vectors, more poral patterns, such as text classifica - tions, namely binding, bundling, and than a third of the bits of a vector can be tion, biomedical signal processing, permutation, and then stored in associa - flipped by randomness, device varia - multimodal sensor fusion, and distrib - tive memories to be recalled, matched, tions, defects, and noise, and the faulty uted sensors [4]. A key advantage is decomposed, or reasoned about. vector can still be identified with the that the training algorithm in HDC Manipulation and comparison of these correct one, as it is closer to the original works in one or only a few shots: that large patterns results in a bottleneck error-free vector than to any unrelated is, object categories are learned from when implemented on the conventional vector chosen so far, with near cer - one or a few examples, and in a single von Neumann computer architectures. tainty. Therefore, the inherent robust - pass over the training data as opposed On the other hand, the chain of opera - ness and the need for manipulations of

ERCIM NEWS 125 April 2021 29 Special Theme

large patterns stored in memory make In our architecture, shown in Figure 1, References: HDC particularly well suited to an associative memory search is per - [1] P. Kanerva, Hyperdimensional emerging computing paradigms such as formed using a second memristive computing: an introduction to in-memory computing or computational crossbar for in-memory dot-product computing in distributed memory based on emerging nanoscale operations on the encoded output hyper - representation with high- resistive memory or memristive devices vectors from the first crossbar, realising dimensional random vectors. Cogn. [5]. the full functionality of the HDC Comput., 2009. system. Our combination of analog in- [2] P. Kanerva, “Computing with In the past few years, we have been memory computing with CMOS logic High-Dimensional Vectors,” IEEE working towards designing and opti - allows continual functioning of the Design & Test, 2019. mising a complete integrated in-memory memristive crossbars with desired accu - [3] A. Mitrokhin, et al. Learning HDC system in which all the operations racy for a wide range of multiclass clas - sensorimotor control with of HDC are implemented on two planar sification tasks, including language neuromorphic sensors: toward memristive crossbars together with classification, news classification, and hyperdimensional active perception. peripheral digital CMOS circuits. We hand gesture recognition from elec - Science Robotics, 2019. have been devising a way of performing tromyography signals. We verify the [4] A. Rahimi, et al. Efficient hypervector binding entirely within a integrated inference functionality of the biosignal processing using first memristive crossbar using an in- system through large-scale mixed hard - hyperdimensional computing: memory read logic operation and hyper - ware/software experiments, in which network templates for combined vector bundling near the crossbar with hypervectors are encoded in 760,000 learning and classification of ExG CMOS logic. These key operations of hardware phase-change memory signals. Proc. IEEE, 2019. HDC cooperatively encode hypervec - devices performing analog in-memory [5] A. Sebastian, et al. Memory tors with high precision, while elimi - computing. Our experiments achieve devices and applications for in- nating the need to repeatedly program comparable accuracies to the software memory computing. Nature (i.e., write) the memristive devices. baselines and surpass those reported in Nanotechnology, 2020. Unlike previous work, this approach previous work. Furthermore, a complete [6] G. Karunaratne, et al. In-memory matches the limited endurance of mem - system-level design of the in-memory hyperdimensional computing. ristive devices and scales well with HDC architecture synthesized using 65 Nature Electronics, 2020. 10,000-dimensional hypervectors, nm CMOS technology demonstrates a making this work the largest experi - greater than six-fold end-to-end reduc - Please contact: mental demonstration of HDC with tion in energy compared with a dedi - Abbas Rahimi, IBM Research Europe, memristive hardware to date [6]. cated digital CMOS implementation. Säumerstrasse 4, 8803 Rüschlikon, More details can be found in our paper Switzerland published in Nature Electronics [6]. +41 44 724 8303, [email protected]

E = AI 2

by Marco Breiling, Bijoy Kundu (Fraunhofer Institute for Integrated Circuits IIS) and Marc Reichenbach (Friedrich- Alexander-Universität Erlangen-Nürnberg)

How small can we make the energy consumed by an artificial intelligence (AI) algorithm plus associated neuromorphic computing hardware for a given task? That was the theme of a German national competition on AI hardware-acceleration, which aimed to foster disruptive innovation. Twenty-seven academic teams, each made up of one or two partners from universities and research institutes, applied to enter the competition. Two of the eleven teams that were selected to enter were Fraunhofer IIS: ADELIA and Lo3-ML (the latter together with Friedrich-Alexander-University Erlangen-Nürnberg - FAU) [L1]. Finally Lo3-ML was one of the four national winners awarded by the German research minister Anja Karliczek for best energy efficiency.

A national competition, sponsored by crossbar arrays of novel analogue pro - is trained using quantized weights of the German research ministry BMBF cessing units (see Figure 1) and stan - seven levels (quasi-3bit), quantized [L2], challenged teams of researchers to dard SRAM cells for storage, while gain (4 bit) and offset (5 bit) of a batch classify two-minute-long electro-cardio - “Lo3-ML” designed a basically digital normalisation (BN), and a customised gram (ECG) recordings as healthy or accelerator with non-volatile Resistive ReLU activation function (AF). showing signs of atrial fibrillation. The RAM (RRAM) cells, for which ana - Crossbar arrays made of novel analogue two teams involving Fraunhofer IIS ran logue write-read-circuits were devel - processing units (APUs) perform the completely independently of each other oped. primary computations of the NN: the and followed completely different multiply and accumulate operations approaches. “ADELIA” implemented a In the ADELIA project, Fraunhofer IIS (MACs). These APUs contain resistors mixed-signal neural network (NN) developed an analogue NN accelerator and switches plus standard CMOS accelerator, which is primarily made of with a mixed-signal frontend. The NN SRAM cells, which hold the trained

30 ERCIM NEWS 125 April 2021 Figure1:Ananaloguedeeplearningacceleratorusesananaloguecrossbarforenergy-efficientanaloguematrixmultiplication,whichisatthe coreoftheneuralnetworkcalculationforclassifyingtheECGsignals.(Sources:FraunhoferIISandCreativeCommonsCC0). weights after chip start-up. The non- aware training algorithms, as corre - [1] J. Knödtel, M. Fritscher, et al.: “A linear activation circuits drive (in ana - sponding tools were not available at that Model-to-Netlist Pipline for logue) their following NN layer, thus time. Moreover, to cope with this quan - Parametrized Testing of DNN avoiding any conversion overhead from tization, a Binary-Shift Batch Norm Accelerators based on Systolic ADCs and DACs. Hardware-software (BSBN) using only the shifting of Arrays with Multibit Emerging co-design was heavily employed during values was introduced. The design and Memories”, MOCAST Conf., training and circuit implementation of implementation of the chip was done 2020. the NN. While the NN training itera - iteratively. In order to achieve quick [2] M. Fritscher , J. Knödtel, et al.: tively took into account several hard - turn-around cycles, the design space “Simulating large neural networks ware constraints such as quantized exploration for both NN and HW archi - embedding MLC RRAM as weight weights, quantized gains and offsets of tecture was partially automated [1] storage considering device BN, customised AF, and some compo - including, for example, automatic gen - variations”, Latin America nent variations, the NN circuit was gen - eration of a bit-true simulation. Symposium on Circuits and erated automatically from the software System, 2021. model using a tool developed during the The projects were realised by intensive project. Thus a fast analogue CMOS work in small teams over just 15 Please contact: based energy efficient NN accelerator months. Despite operating as com - Marco Breiling was developed that harnesses the par - pletely independent projects, both Fraunhofer Institute for Integrated allel processing of analogue crossbar ended up selecting similar NNs with Circuits IIS, Germany computing without inter layer data con - multiple convolutional layers followed [email protected] verters and without the need for mem - by fully connected layers. The achieved ristor technology. metrics are 450 nJ per inference for ADELIA on a 22 nm Global Foundries- The second project, Lo3-ML, was a col - FDSOI technology and 270 nJ for Lo3- laboration between Fraunhofer IIS and ML on a 130 nm IHP process. ADELIA the chairs for Computer Architecture produced one of the very few analogue and Electronics Engineering at FAU. NN accelerators worldwide that are The project built upon their previous actually ready for chip implementation. work with non-volatile RRAMs [2]. Mixed-signal NN acceleration is gener - These allow the accelerator core of the ally considered as very energy-efficient chip to be powered down while idle and hence promising for the next gener - (which is most of the time) to save ation of ultra-low-power AI accelera - energy, while the pre-processing core is tors. Although the task in the project always awake and collects relatively was ECG analysis, the developed archi - slowly incoming data (e.g., at low sam - tectures can also be used for other appli - pling frequency). Once sufficient data is cations and could be easily extended for present, the accelerator core is woken more complex tasks. The design-flows up and does a quick processing before developed in the two projects will be powering down again. This scheme combined in the future and will serve as saves up to 95% of energy compared to the basis for a highly automated design a conventional architecture, where both space exploration tool jointly for NN cores are always on. An RRAM cell can topology and HW architecture. store binary or multi-level values. A trade-off analysis showed that using ter - Links: nary values both in the RRAM cells and [L1] https://kwz.me/h58 as the weights of a severely quantized [L2] https://kwz.me/h59 NN offered the best energy trade-off. References: However, for such extreme quantization we had to develop dedicated hardware- ERCIM NEWS 125 April 2021 31 Special Theme

Brain-inspired visual-Auditory Integration Yielding Near Optimal Performance – Modelling and Neuromorphic Algorithms

by Timo Oess and Heiko Neumann (Ulm University)

Audition equips us with a 360-degree far-reaching sense to enable rough but fast target detection in the environment. However, it lacks the precision of vision when more precise localisation is required. Integrating signals from both modalities to a multisensory audio-visual signal leads to concise and robust perception of the environment. We present a brain-inspired neuromorphic modelling approach that integrates auditory and visual signals coming from neuromorphic sensors to perform multisensory stimulus localisation in real time.

Multimodal signal integration improves camera). Robust alignment of such dis - TrueNorth [3]. There, they ran in real- perceptual robustness and fault toler - tinct spaces is a key step to exploit the time and controlled a robotic test plat - ance, increasing the information gain information gain provided by multisen - form to solve target-driven orientation compared to merely superimposing sory integration. tasks. input streams. In addition, if one sensor fails, the other can compensate for the In a joint project [L1] between the In contrast to vision with its topo - lost information. In biology, integration Institutes of Applied Cognitive graphic representation of space, audi - of sensory signals from different modal - Psychology and Neural Information tory space is represented tonotopically, ities in different species is a crucial sur - Processing at Ulm University, we inves - with incoming sound signals decom - vival factor and it has been demon - tigated the processing of auditory and posed into their frequency components. strated that such integration is per - visual input streams for object localisa - Horizontal sound source positions formed in a Bayesian optimal fashion tion with the aim of developing biologi - cannot directly be assigned a location However, multimodal integration is not cally inspired neural network models relative to the head direction. This is trivial, since signals from different for multimodal object localisation because this position is computed from modalities typically are represented in based on optimal visual-auditory inte - different cues, caused by (i) the tem - distinct reference frames and one needs gration as seen in humans and animals. poral and intensity difference of the sig - to align these frames to relate inferred In particular, we developed models of nals arriving at the left and right ear and locations to the spatial surround. An neuron populations and their mutual (ii) the frequency modulation induced example is the integration of audio and interactions in different brain areas that by the head and shape of the ears. Thus, visual signals. The auditory space is are involved in visual and auditory to spatially localise a sound source an computed by level or time differences of object localisation. These models were agent needs to process auditory signals binaural input whereas the visual space behaviourally validated and, in a subse - over a cascade of several stages. We is derived from positions related to the quent step, implemented on IBM’s modelled these stages utilising biologi - sensory receptor surfaces (retina, brain-inspired neurosynaptic chip cally plausible neural mechanisms that

Visual Auditory Visual Auditory Cortical Cortical Cortical Cortical Input Input Input Input

Cv Ca

core #4 core #5

core #1

  core #6 r core #6 core #2 D 

 
 D  D S1v S1a  q q S2v S2a  q q    

Neuromorphic D

ppool +   

core #6 

 

 +

D  D

D D

+ sen D D  p +

Implementation  
 core #6 
   D qm core #2 core #3 Sv Sa

Visual Auditory Visual Auditory Sensory Sensory Sensory Sensory Input Input Input Input Figure1:ModelArchitectures.Leftpanel,architectureofthemultisensoryintegrationmodel.Bluelinesindicateexcitatoryconnections,green linesindicatemodulatoryconnections,andredlinesindicateinhibitoryconnections.Greenboxismodulatorycorticalsignalthatiselaboratedin greybox.Filledcirclesrepresentmodelneurons,hollowcirclesindicateinputstothemodel.Lettersindicatethenameofneuronsandinputs. Rightpanel,arrangementofthemodelarchitectureontheTrueNorthneurosynpaticchip.Notethesimilarhierarchyofthearchitectureswith differencesonlyinthefinestructure.Adaptedfrom[2],CreativeCommonslicense4.0(https://creativecommons.org/licenses/by/4.0/)

32 ERCIM NEWS 125 April 2021 facilitate subsequent deployment on well as feedback from cortex [2]. Such bio-inspired gamma-tone filters) for neuromorphic hardware. In particular, feedback projections facilitate multi - energy efficient processing of audio and we constructed a model for binaural sensory integration (MSI) responses visual signals, respectively, and a neu - integration of sound signals encoding and lead to commonly observed proper - romorphic processing unit. We defined the interaural level difference of signals ties like inverse effectiveness, within- an event-based processing pipeline from the left and right ear [1]. The modality suppression, and the spatial from sensory perception up to stages of model incorporates synaptic adaptation principle of neural activity in multisen - subcortical multisensory integration. to dynamically optimise estimations of sory neurons. In our model, these prop - Performance evaluation for real world sound sources in the horizontal plane. erties emerge from network dynamics inputs demonstrate that the neural The elevation of the source in the ver - without specific receptive field tuning. framework runs in real time and is tical plane is separately estimated by a A sketch of the functionality of the cir - robust against interferences making it model that demonstrates performance cuit is shown in Figure 1. We have fur - suitable for robotic applications with enhancement by binaural signal integra - ther investigated how multimodal sig - low-energy consumption. tion. These estimates jointly define a 2D nals are integrated and how cortical map of auditory space. This map needs modulation signals affect it. Our model - Further evaluations of this platform are to be calibrated to spatial positions in ling investigations demonstrate that planned, that comprise investigating the the environment of the agent. Such a near-optimal Bayesian integration of ability of the model to flexibly react to calibration is accomplished by learning visual and auditory signals can be altered sensory inputs by, e.g., modi - a registered auditory position map accomplished with a significant contri - fying the position of the microphones. which is guided by visual representa - bution by active cortical feedback pro - Attention mechanisms will additionally tions in retinal coordinates. Connection jections. In conclusion, the results shed enable a selective enhancement of mul - weights are adapted by correlated activ - new light how recurrent feedback pro - tisensory signals during active search. ities where learning is triggered by a cessing supports near-optimal percep - Deeper understanding of such princi - third factor leading to robust map align - tual inference in cue integration by ples and testing their function and ment but is sufficiently flexible to adapt adaptively enhancing the coherence of behaviour on robots provides the basis to altered sensory inputs. representations. for developing advanced systems to self-organise stable orientation, locali - Such calibrated internal world represen - Having analysed the functional proper - sation, and recognition performance. tations establish accurate and stable ties of our models for audio object multimodal object representations. We localisation and multisensory integra - Link: modelled multisensory integration neu - tion, we deployed the neural models on [L1] https://www.uni- rons that receive excitatory and a neuromorphic robotic platform ulm.de/in/neuroinformatik/forschung/s inhibitory inputs from unimodal audi - (Figure 2). The platform is equipped chwerpunkte/va-morph/ tory and visual neurons, respectively, as with bio-inspired sensors (DVS and References: [1] T. Oess, M. O. Ernst, H. Neumann: “Computational principles of neural adaptation for binaural signal integration”, PLoS Comput Biol (2020) 16(7): e1008020. https://doi.org/10.1371/journal.pcbi .1008020 [2] T. Oess, et al.: “From near-optimal Bayesian integration to neuromorphic hardware: a neural network model of multisensory integration”, Front Neurorobot (2020) 14:29. doi: 10.3389/fnbot.2020.00029 [3] P. A. Merolla, et al., “A million spiking-neuron integrated circuit with a scalable communication network and interface”, Science, vol. 345, no. 6197, pp. 668-673, Aug 2014.

Please contact: Timo Oess, Heiko Neumann Dept. Applied Cognitive Psychology; Figure2:Lefttop,eDVSinmounted,vibrationalframedrivenbyvibrationmotorstoinduce Inst. for Neural Information microsaccadesfordetectionofstaticobjects.Leftbottom,leftsideof3Dprintedheadwith Processing, Ulm University, Germany eDVSframeandmicrophoneinearchannel.Rightside,completeheadmountedonrobotic {timo.oess | heiko.neumann}@uni- platform. ulm.de

ERCIM NEWS 125 April 2021 33 Special Theme

Touch in Robots: A Neuromorphic Approach

by Ella Janotte (Italian Institute of Technology), Michele Mastella, Elisabetta Chicca (University of Groningen) and Chiara Bartolozzi (Italian Institute of Technology)

In nature, touch is a fundamental sense. This should also be true for robots and prosthetic devices. In this project we aim to emulate the biological principles of tactile sensing and to apply it to artificial autonomous systems.

Babies are born with a grasping reflex, Motivation tials, or spikes) upon contact. This can triggered when something touches E-skins must cover large surfaces while be applied to an artificial e-skin, imple - their palms. Thanks to this reflex, they achieving high spatial resolution and menting neuromorphic, or event-driven, are able to hold on to fingers and, later enabling the detection of wide band - sensors’ readout. on, manually explore objects and their width stimuli, resulting in the genera - surroundings. This simple fact shows tion of a large data stream. In the H2020 Like neuromorphic vision sensors, the the importance of biological touch for NeuTouch [L1] project, we draw inspi - signal is sampled individually and asyn - the understanding of the environment. ration from the solutions adopted by chronously, at the detection of a change However, artificial touch is less promi - human skin. in the sensing element's analogue value. nent than vision: even tasks such as Initially, the encoding strategy can be manipulation, which require tactile Coupled with non-uniform spatial sam - based on emitting an event (a digital information for slip detection, grip pling (denser at the fingertips and voltage pulse) when the measured strength modulation and active explo - sparser on the body), tactile information signal changes by a given amount with ration, are widely dominated by vision- can be sampled in an event-driven way, respect to the value at the previous based algorithms. There are many rea - i.e., upon contact, or upon the detection event. We will then study more sophisti - sons for the underrepresentation of tac - of a change in contact. This reduces the cated encoding based on local circuits tile sensing, starting from the chal - amount of data to be processed and, if that emulate the slow and fast adaptive lenges posed by the physical integra - merged with on-chip spiking neural net - afferents. Events are transmitted off- tion of robust tactile sensing technolo - works for processing, supports the chip asynchronously, via AER-protocol, gies in robots. Here, we focus on the development of efficient tactile systems identifying the sensing element that problem of the large amount of data for robotics and prosthetics. observed the change. In this representa - generated by e-skin systems that tion, time represents itself and the tem - strongly limits their application on Neuromorphic sensors poral event pattern contains the stim - autonomous agents which require low Mechanoreceptors of hairless human ulus information. Thus, the sensor power and data-efficient sensors. A skin can be roughly divided into two remains idle in periods of no change, promising solution is the use of event- groups: slowly and rapidly adapting. avoiding the production of redundant driven e-skin and on-chip spiking Slowly adapting afferents encode stim - data, while not being limited by a fixed neural networks for local pre-pro - ulus intensity while rapidly adapting sampling rate if changes happen fast. cessing of the tactile signal [1]. ones respond to changes in intensity. In both cases, tactile afferents generate a We aim to exploit the advantages of series of digital pulses (action poten - event-driven sensing to create a neuro -

Figure1:Agraphicalrepresentationofthedesiredoutcomeofourproject.Therealisedarchitecturetakesinformationfrombiologicallyinspired sensors,interfacingwiththeenvironment.Theoutcomingdataaretranslatedintospikesusingevent-drivencircuitsandprovideinputtodifferent partsintheelectronicchip.Thesedifferentpartsareresponsibleforanalysingtheincomingspikesanddeliveringinformationabout environmentalpropertiesofobjects.Theresponsesarethenusedtogenerateanapproximationaboutwhatishappeninginthesurroundingsand impactthereactionoftheautonomousagent.

34 ERCIM NEWS 125 April 2021 morphic e-skin that produces a sparse nition of textures, which can be done by These two novel paradigms have sev - spiking output, to solve the engineering recreating a spiking neural network that eral advantages that will result in a problem of data bandwidth for robotic senses frequencies [3]. A novel archi - structure capable of exploring the envi - e-skin. The signal can be then processed tecture, composed only of neurons and ronment with bioinspired and efficient by traditional perception modules. synapses organised in a recurrent architectures. This combined approach However, the spike-based encoding fashion, can spot these frequencies and can greatly enhance the world of calls for the implementation of spiking signal the texture of a given material. autonomous agents. neural networks for extracting informa - tion. These networks can be built in neuro - Link: morphic mixed-mode subthreshold [L1] https://neutouch.eu/ Spiking neural networks CMOS technology, to emulate SNNs The spiking asynchronous nature of using the same technological processes References: neuromorphic tactile encoding paves of traditional chips, but exploiting an [1] C. Bartolozzi, L. Natale, L., Nori, the way to the use of spiking neural net - extremely low-power region of the tran - G. Metta: “Robots with a sense of works (SNNs) to infer information sistor’s behaviour. By embedding the touch”, Nature Materials 15, about the tactile stimulus. SNNs use networks directly on silicon using this 921–925 (2016). neuron and synapse models that more strategy, we aim for low power con - [2] A. Dabbous, et. al.: “Artificial Bio- closely match the behaviour of biology, sumption, enabled by the neuron’s inspired Tactile Receptive Fields for using spatio-temporal sequences of ability to consume energy only when Edge Orientation Classification”, spike to encode and decode informa - active and to interface directly with ISCAS (2021) [in press]. tion. Synaptic strength and the connec - event-driven data. [3] M. Mastella, E. Chicca: “A tivity of the networks are shaped by Hardware-friendly Neuromorphic experience, through learning. Conclusion Spiking Neural Network for Our goal is to equip autonomous Frequency Detection and Fine Examples of neuromorphic tactile sys - agents, such as robots or prosthesis, Texture Decoding”, ISCAS (2021) tems that couple event-driven sensing with the sense of touch, using the neu - [in press]. with SNN have been developed for ori - romorphic approach both for sensing entation detection, where the coincident and processing. We will employ event- Please contact: activations of several event-driven sen - driven sensors to capture temporal Ella Janotte, Italian Institute of sors is used to understand the angle at information in stimuli and to encode it Technology, iCub facility, Genoa, Italy. which a bar, pressed on the skin, is tilted in spikes and spiking neural networks [email protected] [2]. Different sensors’ outputs are to process data in real time, with low joined together and connected to a power consumption. The network will Michele Mastella, BICS Lab, Zernike neuron, when the neuron spikes it sig - be implemented on a silicon tech - Inst Adv Mat, University of nals that those sensors were active nology, delivering the first neuromor - Groningen, Netherlands together. Another example is the recog - phic chip for touch. [email protected] uncovering Neuronal Learning Principles through Artificial Evolution by Henrik D. Mettler (University of Bern), Virginie Sabado (University of Bern), Walter Senn (University of Bern), Mihai A. Petrovici (University of Bern and Heidelberg University) and Jakob Jordan (University of Bern)

Despite years of progress, we still lack a complete understanding of learning and memory. We leverage optimisation algorithms inspired by natural evolution to discover phenomenological models of brain plasticity and thus uncover clues to the underlying computational principles. We hope this accelerates progress towards deep insights into information processing in biological systems with immanent potential for the development of powerful artificial learning machines.

What is the most powerful computing intertwined cells to perform the of humans and other animals. device in your home? Maybe your required computations? Neurons, However, as physics has long known, laptop or smartphone spring to mind, capable of exchanging signals via the nature of matter is determined by or possibly your new home automation short electrical pulses called “action the interaction of its constituents. For system. Think again! With a power potentials” or simply “spikes”, are the neural networks, both biological and consumption of just 10W, our brains main carriers and processors of infor - artificial, the interaction between neu - can extract complex information from mation in the brain. It is the organised rons is mediated by synapses, and it is a high-throughput stream of sensory activity of several billions of neurons their evolution over time that allows inputs like no other known system. But arranged in intricate networks, that sophisticated behaviour to be learned what enables this squishy mass of underlies the sophisticated behaviour in the first place.

ERCIM NEWS 125 April 2021 35 Special Theme

Figure1:Schematicoverviewofourevolutionaryalgorithm.Fromaninitialpopulationofsynapticplasticityrules(g,h),newsolutions (offspring,g',h')arecreatedbymutations.Eachruleisthenevaluatedusingaparticularnetworkarchitectureonapredefinedtask,resultingina

fitnessvalue(F g,F g',F h,F h').High-scoringrulesareselectedtobecomethenewparentpopulationandtheprocessisrepeateduntilaplasticity rulereachesatargetfitness.

Synaptic plasticity describes how and guided generalisation. Furthermore, this ties of the dataset to evolve plasticity why changes in synaptic strengths take interpretability allows us to extract the rules that learn faster than some of their place. Early work on uncovering gen - key interactions between biophysical more general, manually derived coun - eral principles governing synaptic plas - variables giving rise to plasticity. Such terparts. ticity goes back to the 1950s, with one insights provide hints about the under - of the most well-publicized results lying biophysical processes and also Since our approach requires a large being often summarised by the suggest new approaches for experi - number of neuronal network simula - mnemonic “what fires together, wires mental neuroscience (Figure 2). tions, we make use of modern HPC together”. However, this purely correla - infrastructure, such as Piz Daint at the tion-driven learning is but one aspect of Two of our recent manuscripts have Swiss National Supercomputing Centre a much richer repertoire of dynamics highlighted the potential of our (CSCS) as well as high-performance espoused by biological synapses. More recent theories of synaptic plasticity have therefore incorporated additional external signals like errors or rewards. These theories connect the systems- level perspective (“what should the system do?”) with the network level (“which dynamics should govern neu - ronal and synaptic quantities?”). Unfortunately, the design of plasticity rules remains a laborious, manual process and the set of possible rules is large, as the continuous development of new models suggests.

In the Neuro-TMA group at the Department of Physiology, University Figure2:Symbioticinteractionbetweenexperimentalandtheoretical/computational of Bern we are working on supporting neuroscience.Experimentalneuroscientistsprovideobservationsaboutsingleneuronand this manual process with powerful auto - networkbehaviours.Theoreticalneuroscientistsdevelopmodelstoexplainthedataanddevelop mated search methods. We leverage experimentallytestablehypotheses,forexampleaboutthetimeevolutionofneuronalfiring modern evolutionary algorithms to dis - ratesduetoongoingsynapticplasticity. cover various suitable plasticity models that allow a simulated neuronal network architecture to solve synthetic tasks evolving-to-learn (E2L) approach by software for the simulation of neuronal from a specific family, for example to applying it to typical learning scenarios networks [3] and from the Scientific navigate towards a goal position in an in both spiking and rate-based neuronal Python ecosystem. To support the spe - artificial two-dimensional environment. network models. In [1], we discovered cific needs of our research we have In particular, we use genetic program - previously unknown mechanisms for developed an open-source pure-Python ming, an algorithm for searching learning efficiently from rewards, library for genetic programming [L1]. through mathematical expressions recovered efficient gradient-descent We believe that the open nature of such loosely inspired by natural evolution methods for learning from errors, and community codes holds significant (Figure 1), to generate human-inter - uncovered various functionally equiva - potential to accelerate scientific pretable models. This assures our dis - lent spike-timing-dependent-plasticity progress in the computational sciences. coveries are amenable to intuitive rules with tuned homeostatic mecha - understanding, fundamental for suc - nisms. In [2], we demonstrated how In the future, we will explore the poten - cessful communication and human- E2L can incorporate statistical proper - tial of neuromorphic systems, dedicated

36 ERCIM NEWS 125 April 2021 hardware for the accelerated simulation This research has received funding from [2] H.D. Mettler et al.: “Evolving of neuronal network models. To this the European Union’s Horizon 2020 Neuronal Plasticity Rules using end, we collaborate closely with hard - Framework Programme for Research Cartesian Genetic Programming”, ware experts at the Universities of and Innovation under the Specific Grant 202, arXiv: cs.NE/2102.04312. Heidelberg, Manchester and Sussex. Agreement No. 945539 (Human Brain [3] J. Jordan, et al.: “Extremely Project SGA3). We would like to thank Scalable Spiking Neuronal In summary, our E2L approach repre - Maximilian Schmidt for our continuous Network Simulation Code: From sents a powerful addition to the neuro - collaboration and the Manfred Stärk Laptops to Exascale Computers, scientist’s toolbox. By accelerating the foundation for ongoing support. Frontiers in Neuroinformatics, 12, design of mechanistic models of 2018, synaptic plasticity, it will contribute not Links: doi:10.3389/fninf.2018.00002 only new and computationally powerful [L1] https://kwz.me/h5f learning rules, but, importantly, also [L2] https://kwz.me/h5S Please contact: experimentally testable hypotheses for [L3] (Graphics) www.irasutoya.com Henrik D. Mettler, synaptic plasticity in biological neu - NeuroTMA group, Department of ronal networks. This effectual loop References: Physiology, University of Bern between theory and experiments will [1] J. Jordan, et al.: “Evolving to learn: [email protected] hopefully go a long way towards discovering interpretable plasticity unlocking the mysteries of learning and rules for spiking networks”, 2020. memory in healthy and diseased brains. arXiv:q-bio.NC/2005.14149

What Neurons do – and don’t do by Martin Nilsson (RISE Research Institutes of Sweden) and Henrik Jörntell (Lund University, Department of Experimental Medical Science)

Biology-inspired computing is often based on spiking networks, but can we improve efficiency by going to higher levels of abstraction? To do this, we need to explain the precise meaning of the spike trains that biological neurons use for mutual communication. In a cooperation between RISE and Lund University, we found a spectacular match between a mechanistic, theoretical model having only three parameters on the one hand, and in vivo neuron recordings on the other, providing a clear picture of exactly what biological neurons “do”, i.e., communicate to each other.

Most neuron models are empirical or early stage, that the neuron low-pass fil - the dendrites; and an axon initial seg - phenomenological because this allows a ters input heavily, so clearly, the cause ment compartment consisting of the ini - comfortable match with experimental of the high-frequency component of the tial part of the axon. One way to data. If nothing else works, additional interspike interval variability could not describe the model is as trisecting the parameters can be added to the model be the input, but was to be found locally. classical Hodgkin-Huxley model by until it fits data sufficiently well. The The breakthrough came with the mathe - inserting two axial resistors and disadvantage of this approach is that an matical solution of the long-standing observing the stochastic behaviour of empirical model cannot explain the first-passage time problem for sto - ion channels in the proximal compart - neuron – we cannot escape the uneasi - chastic processes with moving bound - ment. ness of perhaps having missed some aries [1]. This method enabled us to important hidden feature of the neuron. solve the model equations in an instant From the theoretical model we could Proper explanation requires a mecha - and allowed us to correct and perfect the compute the theoretical probability dis - nistic model, which is instead based on model using a large amount of experi - tribution of the interspike intervals the neuron’s underlying biophysical mental data. (ISIs). By comparing this with the ISI mechanisms. However, it is hard to find histograms (and better, the kernel den - a mechanistic model at an appropriate We eventually found that the neuron’s sity estimators) of long (1,000–100,000 level of detail that matches data well. spiking can be accurately characterised ISIs) in vivo recordings, the accuracy What we ultimately want, of course, is by a simple mechanistic model using was consistently surprisingly high to find a simple mechanistic model that only three free parameters. Crucially, (Figure 1c). Using the matching to com - matches the data as well as any empir - we found that the neuron model neces - pute model parameters as an inverse ical model. sarily requires three compartments and problem, we found that the error was must be stochastic. The division into within a factor two from the Cramér- We struggled for considerable time to three compartments is shown in Figure Rao lower bound for all recordings. find such a mechanistic model of the 1b, having a distal compartment con - cerebellar Purkinje neuron (Figure 1a), sisting of the dendrite portions far from It seems that the distal compartment is which we use as a model neuron system. soma; a proximal compartment con - responsible for integrating the input; the Biological experiments revealed, at an sisting of soma and nearby portions of proximal compartment generates a

ERCIM NEWS 125 April 2021 37 Special Theme

ramp and samples it, and the axon initial The major limitation is that we assume to [L1] for a popular description segment compartment detects a stationary input. This is by design, without formulas. threshold passage and generates the because we want to eliminate uncon - spike. trollable error sources such as cortical From a technical point of view, it is input. In the cerebellum, this can be interesting to note that the soft-thresh - We have concluded that the neuron achieved experimentally by decerebra - olding function is nearly identical to the function appears rather straightforward, tion. However, by observing the distal rectifying linear unit (ReLU), and even and that this indicates that there is compartment’s low-pass filtering prop - more so to the exponential, or smooth potential to proceed beyond the spiking erties, it is straightforward to generalize soft-thresholding function that has level towards higher levels of abstrac - the model to accept non-stationary input recently received much attention in the tion, even for biological neurons. The using a pseudo-stationary approach. machine learning field. fundamentally inherent stochasticity of the neuron is unavoidable, and this must So, what do the neurons do, then? In The next step is to investigate the impli - be taken into account, but there is no brief, it turns out that the Purkinje neu - cations for ensembles of neurons. Is it need to worry excessively about hidden rons first soft-threshold the internal possible to formulate an abstraction yet undiscovered neuron features that potential, and then encode it using pulse which treats such assemblies of neurons would disrupt our view of what neurons frequency modulation, dithered by as a single unit without considering are capable of. The reason is the nearly channel noise to reduce distortion. And each neuron independently? perfect match between model and data; this is it! One of the three free parame - the match cannot be significantly ters is the input, and the other two corre - This research was partially funded by improved even if the model is elabo - spond to the soft-threshold function’s the European Union FP6 IST Cognitive rated, which we show using the Cramér- gain (slope) and offset (bias), respec - Systems Initiative research project Rao lower bound. tively. Please refer to [2] for details, or SENSOPAC, “SENSOrimotor struc - turing of Perception and Action for emerging Cognition”, and the FP7 ICT Cognitive Systems and Robotics research project THE, “The Hand Embodied”, under grant agreements 028056 and 248587, respectively.

Link: [L1] https://www.drnil.com/#neurons-doing- what (retrieved 2021-03-16)

References: [1] M. Nilsson: “The moving- eigenvalue method: Hitting time for Itô processes and moving boundaries”, Journal of Physics A: Mathematical and Theoretical (Open Access), October 2020. DOI: 10.1088/1751-8121/ab9c59 [2] M. Nilsson and H. Jörntell: “Channel current fluctuations conclusively explain neuronal encoding of internal potential into spike trains”, Physical Review E (Open Access), February 2021. DOI: 10.1103/PhysRevE.103.022407

Please contact: Martin Nilsson, RISE Research Institutes of Sweden, [email protected]

Figure1:(a)ConfocalphotomicrographofPurkinjeneuron.(b)Proposeddivisionofneuron intothreecompartments.(c)Exampleofthematchbetweentheoreticalmodel(redsolidtrace) andexperimentalinterspike-intervalhistogram(bluebars;blackdashedtraceforkerneldensity estimator).Imagecredits:CCBY4.0[2]. 38 ERCIM NEWS 125 April 2021 NEuROTECH - A European Community of Experts on Neuromorphic Technologies by Melika Payvand, Elisa Donati and Giacomo Indiveri (University of Zurich)

Neuromorphic Computing Technology (NCT) is becoming a reality in Europe thanks to a coordinated effort to unite the EU researchers and stakeholders interested in neuroscience, artificial intelligence, and nanoscale technologies.

Artificial intelligence (AI) has been NEUROTECH is an EU coordination include monthly educational activities, achieving impressive results in a wide and support action that aims to generate where experts are invited to answer a range of tasks. Progress in AI is pro - mutual awareness among organisations, specific question on NCTs, e.g., sup - ducing ever more complex and powerful national networks, projects, and initia - porting technologies for neuromorphic algorithms. However, the process of tives in neuromorphic computing and processing, event-driven sensors, and training and executing these algorithms engineering, and to promote the processors, and circuits for online on standard computing technologies exchange of tools, solutions, ideas, and, learning. Every event is recorded and consumes vast amounts of energy and is where possible, IP among all EU stake - uploaded in an online channel, which not sustainable in the long term. A prom - holders. The participating institutions acts as a collection of introductory ising approach to building a new genera - are the University of Zurich material for students and researchers tion of AI computing technologies that (Switzerland, coordinator), University who want to get involved in neuromor - can dramatically reduce power con - of Manchester (UK), Heidelberg phic computing. The project is com - sumption is “neuromorphic computing University (Germany), University of posed of four subgroups: Industry, and engineering”. This ground-breaking Groningen (Netherlands), Italian Bridge, Science, and Ethics, which aim approach draws inspiration from biolog - Institute of Technology (IIT, Italy), to create a connection with industry and ical neural processing systems. Today’s University of Bordeaux (France), other communities, and to define guide - neuromorphic research community in University of Hertfordshire (UK), lines for ethics in neuromorphic com - Europe is already leading the state of the THALES SA (France), IBM Research puting and engineering. Each work - art. The NEUROTECH project has been GmbH (Switzerland), Consiglio group organises periodic webinars and extremely successful in fostering this Nazionale Delle Ricerche (CNR, Italy), panel discussions to increase awareness community, promoting its growth, and Interuniversitair Micro-Electronica within the community and with a large engaging with stakeholders to enable the Centrum vzw (IMEC, Belgium) and audience. With the exception of an ini - uptake of this technology in industry Commissariat a l’Energie Automatique tial forum, held in Leuven, Belgium, in globally. The NEUROTECH services, et aux Energies Alternatives (CEA 2019, all these events are currently webinars, forums, educational and LETI, France). being held online. public outreach activities are attracting interest from both research institutions The project is a coordination action that The NEUROTECH web portal [L1] and small, medium, and large enter - organises a variety of events around the aims to create a web presence for the prises worldwide. topic of NCTs. The various events project that can be used to make both

ERCIM NEWS 125 April 2021 39 Special Theme

the NCT community and potential part - and prospects. Dedicated discussions that are scheduled regularly. For ners outside the core community aware were held on enabling SME’s to step example, the Science workgroup event of the project and its plans. In addition, into neuromorphic computing and AI as is planned on April 27th for bringing the web portal collects useful resources well as on AI induced ethical questions. together the coordinators of the recently for the community, such as datasets, ref - The forum attracted more than 1000 accepted neuromorphic EU projects to erences (papers, books, courses), and views, with 150 attendees at all times answer the question How novel tech - deliverables. The web portal also pro - for the entire day from all over the nologies can boost neuromorphic com - vides a Roadmap [1] that defines neuro - world. puting? A view from European project morphic computing and its next steps, a consortia. state-of-the-art document that collects The project started in November 2018 knowledge about the latest break - and will continue until the end of 2022. Link: throughs in NCTs, and a cartography Future activities of the project revolve L1] http://neurotechai.eu document that collects links from dif - around the growth of the community ferent national and international net - and bridge between fields that can influ - Reference: works and organisations in Europe to ence and benefit from NCT. These [1] E. Donati et al.: “Neuromorphic facilitate the growth of the NEU - activities include, but are not limited to, technology in Europe : Brain-inspired ROTECH network. It also provides industrial panel discussions about NCT technologies are advancing apace information about upcoming events – opportunities and challenges, scientific across Europe and are poised to help both organised and supported by the discussions about new discoveries and accelerate the AI revolution”, The project and other relevant events world - the development of materials and algo - Innovation Platform, 2020 wide, and provides a feature for sub - rithms that empower NCT, and discus - scribing to the NEUROTECH mailing sions on the ethical aspects of creating Please contact: list through which relevant news about NCT and its applications. Melika Payvand, Elisa Donati, the NCT is announced. NEUROTECH Giacomo Indiveri, University of Forum II that was held online on March The upcoming events include the edu - Zurich, Switzerland 15th 2021 provided an overview on the cational events that happen monthly, +41 44 635 3024 opportunities of AI, technology trends, and the different workgroup activities (melika, elisa, giacomo)@ini.uzh.ch

NeuroAgents – Autonomous Intelligent Agents that Interact with the Environment in Real Time

by Giacomo Indiveri (University of Zurich and ETH Zurich)

Artificial intelligence systems might beat you in a game of Go, but they still have serious shortcomings when they are required to interact with the real world. The NeuroAgents project is developing autonomous intelligent agents that can express cognitive abilities while interacting with the environment.

Tremendous progress is being made learning and neural computation with brains for building a disruptive neuro - globally in machine learning, artificial the latest developments in microelec - morphic computing technology, and (ii) intelligence (AI), and deep learning. In tronic and emerging memory technolo - to build electronic signal processing parallel, especially in Europe, there is a gies. We are designing autonomous sys - systems that use the same physics of strong convergence of AI algorithms, tems that can express robust cognitive computation used by biological neurons neuroscience basic research and tech - behaviour while interacting with the to better understand how animal brains nology development. environment, through the physics of compute. their computing substrate. In particular, However, despite this remarkable the project is making inroads in under - The project started in September 2017 progress, artificial systems are still standing how to build “neuromorphic and is currently entering its final stages unable to compete with biological sys - cognitive agents”, i.e., fully fledged of research. The studies of neural com - tems in tasks that involve processing artificial systems that can carry out putation and AI algorithms (the “mind” sensory data acquired in real time in complex tasks in uncontrolled environ - of the project) paralleled the design of a complex settings and interacting with ments, with low-level adaptive sensory new neuromorphic processor chip (the the environment in a closed-loop setup. processing abilities and high-level goal- “brain” of the project), which has been The NeuroAgents project is bridging oriented decision-making skills. recently delivered and is now being this gap with basic and applied research tested. Meanwhile the “mind” algo - in “neuromorphic” computing and engi - The project has two ambitious, tightly rithms and architectures were mapped neering. Our team is combining the interlinked goals: (i) to study the princi - on older generation neuromorphic recent advancements in machine ples of computation used by animal processors produced in the previous

40 ERCIM NEWS 125 April 2021 NeuroP ERC project, and interfaced phic sensors and motorised platforms References: with robotic actuators (the “body” part (e.g., in active vision stereo setups) [3]. [1] G. Indiveri and Y. Sandamirskaya: of this project). “The importance of space and Future activities of the project involve time for signal processing in The techniques employed combine the connecting all the pieces together (i.e., neuromorphic agents: the challenge study of neural dynamics in recurrent mind, brain, and body), and applying of developing low-power, networks of neurons measured from the results and the know-how gained to autonomous agents that interact neuroscience experiments, with the practical “edge-computing” applica - with the environment”, IEEE simulation of spiking neural processing tions, i.e., applications in which the Signal Processing Magazine 36.6 architectures on computers, and with agent needs to respond to the data that is (2019): 16-28. the design of mixed-signal being sensed in real-time, with low [2] A. Rubino, et al.: “Ultra-Low- analogue/digital electronic circuits that power and without having to resort to Power FDSOI Neural Circuits for emulate the properties of real neurons cloud-based computing resources. Extreme-Edge Neuromorphic and synapses [1,2]. The theoretical neu - Intelligence”, IEEE Transactions roscience studies focus on perception, The final goal of developing brain- on Circuits and Systems I: Regular memory storage, decision-making and inspired sensing, processing, and cogni - Papers, 2020. motor planning. The software simula - tive architectures, all implemented with [3] N. Risi, et al: “A spike-based tions take the high-level results spiking neural network chips, is indeed neuromorphic architecture of stereo obtained from the theory and implement within reach and the NeuroAgents vision”, Frontiers in neurorobotics them using software spiking neural net - project promises to produce very inter - 14, (2020): 93. work simulators. The software is esting results. designed to incorporate the “features” Please contact: of the electronic circuits used into the NeuroAgents is an ERC Consolidator Giacomo Indivieri simulated neural networks. These fea - project (No. 724295) hosted at the University of Zurich tures include restrictions such as limited Institute of Neuroinformatics at the [email protected] resolution, noise, or the requirement to University of Zurich and ETH Zurich, clip signals to only positive values (e.g., Switzerland. The project has led to col - because voltages cannot go below the laboration with many EU partners, “ground” reference level and currents including IBM Research Zurich; IIT can only flow in one direction). The Genova, Italy; CEA-LETI Grenoble, microelectronic techniques focus on the France; the University of Groningen, design of spike-based learning and plas - the Netherlands; Newcastle University, ticity mechanisms at multiple time UK; or the SME SynSense AG, Zurich scales, and of asynchronous event- Switzerland. based routing schemes for building large-scale networks of spiking neu - rons. The robotics activities include testing the models and the spiking neural network chips with neuromor -

ERCIM NEWS 125 April 2021 41 4 2 R e s e a r c

European Research and Innovation h

a n d

I n n o v a t i o n • • • • b b l T e a s c [ H i m b h r i T H t p W p a i l H H l o l r o s k h H b H e o h n i e m o L e e i w h g o x r e r v e l u e y u o f h h u o a u u a u u u e

i s e n o a e s e i A t n g c v C i f l E t t c P

1 h

e l r e p y p m t w n p e e r n a n i n i o i m r p D i m m i m s m c r l

m o . w u o i g u e i r s ] a r n o m o e

n r r r m n

. s o e

n a d r c r

g t c

e t o a t

,

a

a

a t s r a H u a e r v r t n i n o e t o i i a a 5

a w i r s o o S a t g b t e i u ) k l a p t

r p t c v

b a n i s n c c o n n r i r v r i o

f i n n e e .

m t . m v s - 0 e n n m w n n u

o y o r n o

h e

l i

o a e p t c e i a

h e s - g t e - g o t r x n - - t r o g p e o c i 0 c e C - t i i m - r s p n a i l - u

h s m f e

r n e e

o

l l c n e

u o t l , i .

a l s s

a n b g a

w a r s t

c

l i i o

w i i R

t

i n

r a t

i e o e d s o m c

i e y , k k t k t n l

a n s i s a i o i i r t i g e l C m a k o

l c

a a t p h

g b g

o s

k n a i a m

m f i

g

e y e e a i s n a a e n e

e e s q m n r c

o l c e t e p i m v i

n

p e o o v w y n t

s

i n e

g s i n s l r

o l e - e n d s a

u

r h h A e A A

d f s

e . n w r

e d l p l e r l r g t

o g e

i

A g g u d b l i 2

b u

b i b m i o i e

m i e i

s i a n A u p s e i , d c f t u

t i a o d n i t e c i I k I I e u t l n e e . a n i

o o u

s i e m i

r h

s b d h s e I l r

h o e

n s h e a l n e i v n s e n d

s r b a u I e i t e e w t u l t

f e x e

t p d t p s e

l o s t m s w t t s b e . m e l l i e -

w

a a h

r e g

d s . i e

o r

n a

t

i

A

, y i o e e

r s

u v r e i r c T t s a b i a p i

l o a b

b a i e m t

t u E s i l r e l s g i h f e s e a p e i w p ( c n r a e u o e d r t o e b s v e c n l I r e a n s t i e e h s l W

o

t r i s

l r a e

a e v

r

g

e a i o

o l

a l e a u

y . l e t a , d

c i

r h e i h s o l k o c t n e p c l l r r

t r c

r l k e y d n t u

n c v e a s o s r p n a 3 b s

e l

r c t

i s u e d r a u a w t u o o p h

e a a n a

c n d o t

t g t n e s

n e i t s g s y

e

a

r a l l n u o C b l e e

m i g b g c n b d h r a u i o e i w r e t

p e i d g o i

e

e g , t r v r

i t c r m s

s s p i i f t o t s n n

n s h o n / d y e e s f n r r i n c a

s

b

n d f y n b a

e o a i t

o , l e

p p o l o y

t E r s

g d t g t i

p

i t g r

a a

l n s s

e e

i

c c

i

c o w n i u

s t n r

n a o o n s

c s s y e e n o b t

n c R

d t f r d o

i

o e

m o k u , v u s

i u s A

r a t m o h w a s d v a s t f o o e

r o s

a s o i g a v e u e u u w s i C e a r t

s b e c r a

o e b

i n l n o

e c e u

p w n o u h e i o r v a t t s

i e t t a m i r t p l i o s u l h

a a m I r o e

m

o s i e u

e d

t t t

l - .

a

t n n s e n M s

o

i n I a y i r c s h a t l b i r

e

r e h p b h o o x

A u , g l s

g a l

d c u n

e i f I e a

i e g r d e

t f

h b

f i

c e

e a a s

t e i e w , f o o t t d ) n r o a o o h g I d e w

e

o c p d d c t a

t a a e

o o n

e

a l r p

t

d a

n t r y

a a e w f v e i n n t t i

a r w

e a

e r t r n u m s d h m a t n g

i

i v e n . i p

n h h w e h

s i l x d x

e w t o d t t h w n t

t

b r o o f t a p i s c l

e e s n g i i t r s

l a h d o l T e a s i l i

. p u u u o e l i v e l

e e f n o b i s e o i u t n r p l e e s

a n n e i

t u l s e

l

m n b d e

i r h

l r r n T m

n t

a t h t e t c s t

o o a p l s l

n s r i b f c s t t t

e t c d t a i i o h d e a s s e

a

i

n n o n u t o s e t e i i o n i a h h v u e

s

i - b s o s c t i

a s n r a r e o

r n l a r h

m o u i a c

p l g b n r t g e s i d i

e e h l t s

i e a o

t a

W y e n n n t n t l n n g

n e a

e t p n e i , l s o r e r , r n a

t i

i t u t i s p f n

e

s m d t h

a

a o t

a v i

v u g

a h

e a s p g

e n d g f s f q e i n s g h o r i

c a m n a 3 m i r v r e

s e r

e t l l

o s e o e s b a n l e s e a

i x n u u o c l t e a i

o a d o c e p o t k t C a a n e x n . c t r d

o t o k e r s v - t i n a o

h t e h

a e u u c d

i w . e a t t w r t c n s

u t i d t s d A c

e

t

m k e o s e e n

h e o c i u l i u s m s i r r d n , r d C l f u e h o m t r e b c a

e o , t

n o f o e k t s

e a o i e u i n n a t o e

t s e u e

o f c e e u i

r w a i

l e v o o n r

o i h

o u s t s r

r l m n r r a

e g l

y m x c c a p t t a r n i

s r n n h h

t n s l a t g

g

o

e s t i w r i n t n i p

c e h t h c s l

n h l p i f e h n b h p

m d i t a e s p

v e l x a y u y n e

s o m a t r d l y

( e h

s . e t t r u e e a c e d u y l

l e v

p u l e w h

o l a i n a , r , i i h s l a i e e e

e

f e r y

c c t t h

i e

t

y t . . r

e a

n i d J a d , m y m r c g a s p a w t e

m e v d i i

y r r r

s c a o t

h s s

u

t a i s r n v c g M i i u

i o e s e e

t u c

. d s

h r , i g e e o c . i u f n n a c p

v h a

n e o t c e

e r o i v e e

d s

s m i c a t f a s e p e

a t e n a e c r o c e p g L l , a i d o k w t r n u u e r

s e n o a i t .

a g . s p A e m

a t n a i u g s r l f

n k o n t r

S i c

p r

s t g l i w m n g e s C l l o t u e

e i s b o

h d o l A e t g a t e e o a v E r

s s n

h d I i u i c

c n b

a h e d i , d r h n t l w v i a n R

n s l o n

c n d c e ,

n t n

i w a t s r

e t o

r e r b C s i o y e

C s w i

n i t e

i e v e i o g o g i

t c i e t , , p n g h a w n r p , n ’

o a e I o e e

o d n p

s n

i i M d g n o r

e s o

d c n u

g l r m n i i

e v a t g o a l A d g l a r

n f

u n y t e a g e

o n

t t

i m h o a m f l

i p h n r e N

l v s n d i e o b

w s

e o h s v i v s v s i t - e t s c c e , d n r y s I g r o g E

a m o n i i . o l

d e o

n d f i

r m

i u o

. e i k d h w o a W y v e t

s

i l u

e n c f r r

d a p i n

q p g

F l e r w i n l a i m f l d n ,

c

u d u t p o C e e S l o x a t d v i s o i s u t e i e g e u

e

r r i u t v n i g

t e i v s i

n w n o s e

n e s s

s a 1 n o p a o i g s e n u i n o l a e e

o e e x i o t l

d l r e

2 i e o u v l k t r

l

l t

a m t h

s l g m u s a i s t l n n p

a n p e o S

i n s f t 5

i w

t r m d

f o t o t t e p s a

n

a h l e p t

a

e r n r

c c o t b l g y h i . f l o y – t a a

e

i

m

i e

i

i n e

e o p c e o

s

g n n

e u t i e l n o b l b g A e

h e t a I i f b c b n s a

i , n G d e n c ,

d o i t e o p r , t n w t s r i 1 t r u y

t l n u o b

a b u s u r t u s e

d .

s r k r l a p o e c e

t , d u

l d

r r v i n

i : s r t l t

f o r d s s l c h

e

m s h e e s

p

r t s y

t l o m o

, w T t u r a

o f e o a a 2 o i o s t n o a s

h i i i n a x e a i t i a f e c a

- m 0 s c l n

o n r n n n u e a k h h c , u f y o o o v i g n v

e i t c e a d r a 2 e s

y i l 1 n g n o g p d h g n d d e e e a s s s 1 ------r r f l l t , . . skills. Human languages are complex yet easily learned by children [2], and we need to emulate that for scalable AI. Semantics is represented as chunk-based knowledge graphs in contrast to computational linguis - tics and deep learning, which use large sta - tistics as a weak surrogate. Human-like AI doesn’t reason with logic or statistical models, but rather with mental models of examples and the use of metaphors and analogies, taking inspiration from human studies by Philip Johnson-Laird [L3, 3].

Humans are good at mimicking each other’s Figure1:Cognitivearchitecturewithmultiplecognitivecircuitsequivalenttoashared behaviour. For example: babies learning to blackboard. smile socially at the age of 6 to 12 weeks. Language involves a similar process of mimicry with shared rules and statistics for generation and understanding. Work on cognitive natural language processing is focusing on the end-to-end communication of meaning, with constrained working memory and incremental processing, avoiding backtracking through concurrent processing of syntax and semantics.

Work on human-like AI is still in its infancy but is already providing fresh insights for building AI systems, combining ideas from multiple disciplines. It is time to give com - puters a human touch! Figure2:Cortico-basalgangliacircuitforcognition. An expanded version of this article is avail - able [L4]. The described work has been sup - instrument where muscle memory is needed to control ported by the Horizon 2020 project Boost 4.0 (big data in your finger placements as thinking explicitly about each smart factories). finger would be far too slow. Links: Zooming in on cognition, we have the following architec - [L1] https://www.w3.org/community/cogai/ ture, which derives from work by John Anderson on ACT-R [L2] http://act-r.psy.cmu.edu/about/ [L2, 1]. The buffers each hold a single chunk, where each [L3] https://www.pnas.org/content/108/50/19862 chunk is equivalent to the concurrent firing pattern of the [L4] https://www.w3.org/2021/Human-like-AI-article- bundle of nerve fibres connecting to a given cortical region. raggett-long.pdf This works in an analogous way to HTTP, with buffers acting [L5] https://www.w3.org/2021/digital-transformation-2021- as HTTP clients and the cortical modules as HTTP servers. 03-17.pdf The rule engine sequentially selects rules matching the buffers and either updates them directly or invokes cortical References: operations that asynchronously update the buffers. [1] J. R. Anderson: “How Can the Human Mind Occur in the Physical Universe?”, Oxford University Press, A lightweight syntax has been devised for chunks as collec - 2007, https://doi.org/ tions of properties for literals or names of other chunks, and 10.1093/acprof:oso/9780195324259.001.0001 equivalent to n-ary relationships in RDF. Chunks provide a [2] W. O’Grady: “How Children Learn Language”, Cam - combination of symbolic and sub-symbolic approaches, with bridge University Press, 2012, graphs + statistics + rules + algorithms. Chunk modules sup - https://doi.org/10.1017/CBO9780511791192 port stochastic recall analogous to web search. Chunks [3] P. Johnson-Laird “How We Reason”, Oxford University enable explainable AI/ML and learning with smaller datasets Press, 2012, using prior knowledge and past experience. https://doi.org/10.1093/acprof:oso/9780199551330.001. 0001 Symbolic AI suffers from the bottleneck caused by reliance on manual knowledge engineering. To overcome this chal - Please contact: lenge, human-like AI will mimic how children learn in the Dave Raggett, W3C/ERCIM classroom and playground. Natural language is key to [email protected] human-agent collaboration, including teaching agents new

ERCIM NEWS 125 April 2021 43 Research and Innovation

Graph-based Management of Neuroscience data: Representation, Integration and Analysis

by Maren Parnas Gulnes (University of Oslo / SINTEF AS), Ahmet Soylu (OsloMet – Oslo Metropolitan University) and Dumitru Roman (SINTEF AS)

Advances in technology have allowed the amount of neuroscience data collected during brain research to increase significantly over the past decade. Neuroscience data is currently spread across a variety of sources, typically provisioned through ad-hoc and non-standard approaches and formats, and it often has no connection to other relevant data sources. This makes it difficult for researchers to understand and use neuroscience and related data. A graph-based approach Figure1:Anoverviewofinitiativesinvestigatedforoverlapwith could make the data more accessible. themurinebasalgangliadataset.

A graph-based approach for representing, analysing, and accessing brain-related data [1] could be used to integrate cific parts of the cell, while the morphologies describe the various disparate data sources and improve the understand - cell's physical structure. The dataset's primary purpose is for ability and usability of neuroscience data. Graph data models researchers to find and compare neuroanatomical informa - and associated graph database management systems provide tion about the basal ganglia brain regions. performance, flexibility, and agility, and open up the possi - bility of using well-established graph analytics solutions; To identify datasets that overlap with the murine basal gan - however, there is limited research on graph-based data repre - glia dataset for integration purposes, we evaluated a set of sentation as a mechanism for the integration, analysis, and related data sources, including repositories, atlases, and pub - reuse of neuroscience data. licly available data, against the following criteria: (i) serves data programmatically; (ii) contains data related to the basal We applied our proposed approach to a unique dataset of ganglia; and (iii) provides data that could be connected to quantitative neuroanatomical data about the murine basal murine basal ganglia. Figure 1 summarises the results of our ganglia – a group of nuclei in the brain essential for pro - investigation; Brain Architecture Management System cessing information related to movement. The murine basal (BAMS) [L1], InterLex [L2], and NeuroMorpho.Org [L3] ganglia dataset consists of quantitative neuroanatomical data matched the specified criteria. about basal ganglia found in healthy rats and mice, collected from more than 200 research papers and data repositories [2]. We designed and implemented a graph model for the murine The dataset contains three distinct information types: quanti - basal ganglia dataset and migrated the data from the rela - tations (counts), distributions, and cell morphologies. The tional database into a NoSQL graph database [3]. Further, we counts and distributions relate to either entire cells or spe - designed and implemented the integration of data from

Figure2:(a)Relationshipbetweenthedatasetanalysesandthesexand(b)thedatasetanalyseswithrelatednodes.

44 ERCIM NEWS 125 April 2021 related neuroscience data sources and the technical solution for graph-based data analysis. To provide web-based access The ICARuS Ontology: to the graph data, we designed and implemented a web appli - cation and API. The main components of our approach were: A General Aviation (i) the common graph model based on a native graph data - base management system; (ii) integration of identified Ontology external datasets through an extract-transform-load process; (iii) graph analytics to analyse the graph data, based on by Joanna Georgiou, Chrysovalantis Christodoulou, existing graph algorithms and visualisation methods; (iv) George Pallis and Marios Dikaiakos (University of Cyprus) web-based data access interface for the data based on the graph model. The database generated in this study consists of A key challenge in the aviation industry is managing 9,539 distinct nodes with 46 distinct node labels, 29,807 dis - aviation data, which are complex and often derived from tinct relationships, and 66 distinct relationship types [L4]. heterogeneous data sources. ICARUS Ontology is a domain-specific ontology that addresses this challenge We conducted exploratory and confirmatory data analyses on by enhancing the semantic description and integration of the data. The former aims to obtain general information the various ICARUS platform assets. about the data and the latter to answer specific questions. For the exploratory analysis, we used community detection algo - The current digital revolution has heavily influenced the rithms to investigate the graph data structure (Label propaga - way we manage data. For the past few years every human tion and Louvain algorithms), a centrality algorithm to find activity, process, and interaction has been digitised, influential nodes in a graph (PageRank), a node similarity resulting in an exponential increase in the amount of data algorithm to compare nodes (available in Neo4j), and graph produced. There is huge potential for useful information to visualisations (ForcedAtlas2) to investigate the general data be extracted from this data, perhaps enabling us to discover structure. A consultation with a neuroscience expert revealed new ways of optimising processes, find innovative solu - that the results include interesting expected and unexpected tions, and improve decision-making. However, managing findings as well as already known findings. For example, enormous amounts of data from numerous, heterogeneous Figure 2 (a) shows that males (largest green node) are studied sources that do not share common schemas or standards far more often than females. In the confirmatory data poses many challenges: data integration and linking remains analysis part, we aimed to find similar analyses based on a a major concern. specific criterion using a node similarity algorithm. For example, we searched for studies (i.e., analyses) that investi - The procedure of integrating and linking data can be expen - gated the same cell type in the same brain region and with the sive and is often underrated, especially for small and same object of interest. Figure 2 (b) presents the studies (in medium-sized enterprises (SMEs), which may lack the orange) in the dataset connected to the specified nodes and required expertise and be unable to invest the necessary time species. The yellow nodes represent the two species in the and resources to understand and share the digital informa - dataset, and the central node in the middle is the cell type tion derived from their operational systems. Usually, data “neurons”. models are created in a manner that can handle information by encoding the structure, format, constraints, and relation - The results and our experience in representing, integrating, ships with real-world entities. and analysing basal ganglia data show that a graph-based approach can be an effective solution, and that the approach The challenge of managing big data is apparent in various should be further considered for management of various industry domains, including the aviation industry. types of neuroscience data. Unfortunately, aviation data providers use very distinct data models that can vary across different dimensions [1], like Links: data encoding format, data field naming, data semantics, [L1] https://bams1.org spatial and temporal resolution, and measurement unit con - [L2] https://scicrunch.org/scicrunch/interlex/dashboard ventions. To enhance the integration of data in a data plat - [L3] http://neuromorpho.org form for the aviation industry, we designed and introduced [L4] https://github.com/marenpg/jupyter_basal_ganglia the ICARUS ontology [2].

References: The ICARUS [L1] platform helps stakeholders that are con - [1] R. Angles and C. Gutiérrez: “Survey of graph database nected to the aviation industry by providing them with a models”, ACM Computing Surveys, vol. 40, no. 1, pp. system to share, collect or exchange datasets, knowledge, 1–39, 2008. and skills. Through these services, the ICARUS project [2] I.E. Bjerke, et al.: “Database of literature derived cellu - seeks to help stakeholders gain better perspectives, optimise lar measurements from the murine basal ganglia”, Sci - operations, and increase customer safety and satisfaction. entific data, vol. 7, no. 1, pp. 1–14, 2020. Additionally, it aims to offer a variety of user-friendly [3] R. Cattell: “Scalable SQL and NoSQL data stores”, Sig - assets, such as datasets, algorithms, usage analytics and mod Record, vol. 39, no. 4, pp. 12–27, 2011. tools. To manage the complex and dynamic evolution of these assets, we sought to develop an ontology to describe Please contact: various information resources that could be integrated and Dumitru Roman, SINTEF AS, Norway managed by the ICARUS platform. Ontologies can act as [email protected] valuable tools to describe and define, in a formal and

ERCIM NEWS 125 April 2021 45 Research and Innovation

Figure1:Ontology andknowledge-base extensionofthe ICARUSplatform architecture.

explicit manner, relationships and constraints, and for key stages: (i) using the ontology to collect aviation and knowledge representation, data integration, and decision health related data based on concepts / entities stored in the making. ontology; (ii) combining the datasets based on their relation - ships as described in the ontology; and (iii) using SPARQL The ICARUS ontology, as is depicted in Figure 1, enables: queries to derive new information and insights from the com - (i) the integration and description of numerous sources of bined datasets. heterogeneous aviation data; (ii) the semantic representation of the metadata generated by the ICARUS platform and the The main innovative aspects of the ICARUS ontology are: knowledge-based storage, which is dynamically updated and (i) the reuse of existing domain ontologies while adding new accessed through queries. More specifically, the ICARUS concepts, definitions, and relations in order to create a new ontology extracts, stores, and represents semantically in the integrated domain ontology for aviation; and (ii) the use of knowledge-based storage all of the platform’s datasets and multiple layers to represent both metadata and aviation-spe - operations. It represents other modules of the ICARUS plat - cific concepts. With this approach, the ontology's domain can form semantically, like deployed algorithms, user interac - also be used for other purposes, for instance, general data tions, service assets, and their popularity. In addition, the marketplaces. ontology keeps the ICARUS data model and knowledge- based storage up-to-date, while facilitating the ongoing inte - The ICARUS ontology can be utilised to perform various gration of new datasets and services. In this way, it also analytic SPARQL queries and extract knowledge. It can be enables searches and queries of various heterogeneous data found in our github repository [L2] and it is intended to be sources that are available on the ICARUS platform. Lastly, used by: the ontology provides an application-programming interface • data providers and consumers, who can be stakeholders to feed valuable knowledge into the ICARUS recommenda - that are directly or indirectly connected in the aviation tion engine algorithms. value chain industry; • people from the IT industry who are supporting the avia - The ontology can be applied in various scenarios, including tion value chain. (e.g., IT companies, web entrepreneurs, the management of epidemic data. Epidemics, such as the and software engineers); COVID-19 pandemic, present a significant challenge for • research organisations and universities studying the avia - health organisations when it comes to as locating, collecting tion ecosystem; and integrating accurate airline and human mobility data • the general public (e.g., passengers); with adequate geographical coverage and resolution. If such • other data marketplaces. challenges are addressed well, this can lead to improved epi - demic forecasts [3] and potentially enable better estimates of ICARUS is a three-and-a-half year project (started 1 anticipated relative losses of revenue under various scenarios January 2018), funded by the European Union’s Horizon by estimating the reduction of passengers in the airline 2020 research and innovation programme. This work is par - mobility network. The ICARUS ontology can semantically tially supported by the EU Commission in terms of combine epidemic and aviation-related data for data ana - ICARUS 780792 (H2020-ICT-2016-2017) project and by lytics and epidemic predictions. These analytics and predic - the Cyprus Research Promotion Foundation in terms of tions can be extracted from the ontology by completing three COMPLEMENTARY/0916/0010 project. The authors are

46 ERCIM NEWS 125 April 2021 solely responsible for the contents of this publication and Blockchains as ledgers for cryptographic currencies cur - can in no way be taken to reflect the views of the European rently exist in a rather special environment, since through Commission. their development and the original stakeholders, they typi - cally do not have to adhere to the same levels of accounta - Links: bility and privacy that traditional IT systems do. This [L1] https://www.icarus2020.aero/ approach may work for a system that exists outside of the tra - [L2] https://github.com/UCY-LINC-LAB/icarus-ontology ditional financial mechanisms, but the scenario is different in the traditional IT landscape, like banking or land registers. In References: these environments, the inherent risks of this more “free” [1] R. M Keller: “Ontologies for aviation data manage - approach will not be accepted by typical end-users, or even ment”, in 2016 IEEE/AIAA 35th Digital Avionics Sys - be illegal. Some experts believe that a major reason for the tems Conference (DASC), pages 1–9. IEEE, 2016. slow acceptance of novel blockchain applications outside the [2] D. Stefanidis, et al.: “The ICARUS Ontology: A general purely speculative area of cryptocurrencies is that trust, aviation ontology developed using a multi-layer accountability and thorough control over the data on the approach”, in Proc. of the 10th International Confer - chain are currently unavailable for blockchain mechanisms ence on Web Intelligence, Mining and Semantics [2]. (WIMS), 2020. [3] C. Nicolaides, et al.: “Hand‐Hygiene Mitigation Strate - One major problem when introducing blockchain-based sys - gies Against Global Disease Spreading through the Air tems is that they are usually treated in the same way as any Transportation Network”, Risk Analysis, vol. 40, no. 4, other form of IT system with respect to security manage - pp. 723–40, 2020. ment. However, they possess several features that differ fun - damentally, depending on the exact type of blockchain (e.g., Please contact: public vs. private, type of consensus mechanism), as well as George Pallis, University of Cyprus, Cyprus the implementation (see Figure 1). One major difference is [email protected] the highly distributed nature of the systems, which is other - wise generally restricted to niche applications or highly unregulated applications like file sharing and bit torrent. Furthermore, unlike most other applications, blockchain- based systems lack a mechanism to remove data from the Security Management chain; on the contrary, most blockchain systems do not sup - port the deletion of data at all. This is in stark contrast to most and the Slow Adoption filesharing applications, where data is rendered inaccessible after a while, depending on the interest of the other partici - of Blockchains pants, and at some point, typically dwindles into obscurity. In blockchain applications on the other hand we must assume by Peter Kieseberg, Simon Tjoa and Herfried Geyer (St. that the attacker might be in possession of any arbitrary Pölten University of Applied Sciences) selection of old versions of the chain, including all its con - tent. This is especially important when considering the issue Blockchains offer a valuable set of tools to provide of access control. When it comes to providing cryptographic resilient distributed solutions for applications that access control, unlike in normal distributed systems it is not require a high level of integrity. Still, apart from being possible to update protection by re-encrypting all data on the used to store cryptocurrency data, blockchains are not chain, as the attacker can always be in the possession of old used in many other areas. This article gives an overview versions. of the major obstacles to the uptake of blockchains, from an information security perspective. Another trust-related problem, for which a solution would be a key to the widespread adoption of blockchains, is the issue Security management is one of the most important parts of of redaction of wrong information entered into the chain. any resilient security strategy for real-life systems. A While there already exist methods for marking this informa - defender must be able to detect incidents, act coherently and tion as invalid, it cannot be removed altogether from the according to best practices, and to learn from attackers in order to develop miti - gation strategies for the future.

Several standards and best practices have been devised; the best known of which is arguably the ISO27k family [1]. Standardisation is a particularly impor - tant process in security management, not only to address technical issues like compatibility with other organisations, but also as an insurance for the decision- makers within those companies, for when large breaches occur. Figure1:Keysecuritymanagementconcernsintheblockchainecosystem. ERCIM NEWS 125 April 2021 47 Research and Innovation

chain. This is especially important when illegal content like child pornography is introduced into a chain, where full and Trick the System: Towards un-recoverable deletion is required. While there are solutions to this problem, they typically do not remove the illegal data understanding Automatic from the chain but merely render it inaccessible. Managing the system as a responsible entity thus can even result in Speech Recognition legal problems. Furthermore, the nature of these distributed systems is highly volatile, with nodes entering and leaving, Systems new nodes being introduced and old ones never operating again. This makes asset management very complex, espe - by Karla Markert (Fraunhofer AISEC) cially since it is unclear to the system provider, which nodes work together, or might belong to the same entities. Automatic speech recognition systems are designed to transcribe audio data to text. The increasing use of such Last, but not least, another key issue from a security manage - technologies makes them an attractive target for ment perspective is heterogeneous systems: Most compa - cyberattacks, for example via “adversarial examples”. nies already run traditional IT systems and won’t change all This article provides a short introduction to adversarial of it to blockchains at once, which means that blockchain- examples in speech recognition and some background based systems need to interface with traditional programs on current challenges in research. that use totally different approaches for deciding when a transaction is valid, how data is stored and access managed. As their name suggests, neural networks are inspired by the Thus, these interfaces would pose an important attack sur - human brain: just like children learn the abstract notion of face, especially since there are currently no standardised “an apple” from a collection of objects referred to as solutions to securely integrate arbitrary blockchains into “apples”, neural networks automatically “learn” underlying complex traditional systems [3]. structures and patterns from data. In classification tasks, this learning process consists of feeding a network with some Security management is a major issue in IT security, as many labelled training data, aiming to find a function to describe practical problems in securing complex environments cannot the relation between data and label. For example, a speech be reduced to pure technical issues. Furthermore, technical recognition system presented with audio files and transcrip - solutions are often rendered useless because it is impossible tions can learn how to transcribe new, previously to introduce them to the system. Security management is thus unprocessed speech data. In recent decades, research has one of the most important aspects in security, and given the paved the way to teaching neural networks how to recognise complexity and special requirements of blockchains, novel faces, classify street signs, and produce texts. However, just approaches are required. as humans can be fooled by optical or acoustic illusions [L0], neural networks can also be tricked to misclassify input data, In the JRC for Blockchain Technologies and Security even when they would be easy to correctly classify for a Management [L1], we tackle these issues and provide best human [1]. practices for adopting blockchain-based systems in addition to traditional IT systems. Our focus is on the issues of trust Here we discuss two popular network architectures for and access control, but we also provide best practices for speech recognition, explain how they mimic human speech dealing with illegal content from the design phase onwards. perception, and how they can be tricked by data manipula - Furthermore, we provide insights into how to adopt stan - tions inaudible to a human listener. We discuss why dards from the ISO27k family as applied to the particularities designing neural networks that are robust to such aberrations of blockchain-based systems. is hard, and which research questions might help us make improvements. Link: [L1] https://research.fhstp.ac.at/projekte/josef-ressel-zen - Speech recognition can be based on different statistical trum-fuer-blockchain-technologien-sicherheitsmanagement models. Nowadays neural networks are a very common approach. Automatic speech recognition (ASR) models can References: be realised end-to-end by one single neural network or in a [1] ISO/IEC 27001:2017. Information Technology - Secu - hybrid fashion. In the latter case, deep neural networks are rity Techniques – Information Security Management combined with hidden Markov models. Systems – Requirements. [2] L. König, et al.: “The Risks of the Blockchain A When training an end-to-end model like Deepspeech [L1] or Review on Current Vulnerabilities and Attacks. J. Inter - Lingvo [L2], the system is only provided with the audio data net Serv. Inf. Secur., 10, 110–127, 2020. and its final transcription. Internally, the audio files are often [3] L. König, et al.: “Comparing Blockchain Standards and pre-processed to “Mel-frequency cepstral coefficients”. In Recommendations. Future Internet, 12(12), p.222, this case, the audio signal is cut into pieces, called frames, 2020. which are in return decomposed into frequency bins approxi - mately summing up to the original input. This pre-processing Please contact: step reflects how the inner ear transmits sounds to the brain. Peter Kieseberg A recurrent neural network then transcribes every frame to a St. Pölten University of Applied Sciences, Austria character without any additional information. Using a spe - [email protected] cific loss function, sequences of frame-wise transcriptions

48 ERCIM NEWS 125 April 2021 Figure1:Adversarialexamplescanfool speechrecognitionsystemswhilebeing imperceptibletothehuman.Thisenablesfar- reachingattacksonallconnecteddevices.

are turned into words. This process mimics our brain under - tems, the transferability of audio adversarial examples standing words from a series of sounds. between different models, the design of detection methods for adversarial examples or of new tools to measure and On the other hand, hybrid models like Kaldi [L3] consists of improve robustness, especially in the language domain. different submodels that are trained separately and require Here, the feature set (audio), the human perception (psychoa - additional expert knowledge. The audio data is pre- coustic), and the learning models (recurrent neural networks) processed in a similar way to above. The acoustic model con - differ from the image domain, on which most previous sists of a neural network that turns these audio features into research has focussed so far. phonemes (provided by an expert). Subsequently, the phonemes are turned into words by a language model that In a way, acoustic illusions and audio adversarial examples also accounts for language-specific information such as are similar: the perception of the human or the neural net - word distributions and probabilities. work, respectively, is fooled. Interestingly, very rarely can the human and the machine both be fooled at the same time Despite all their differences in the mathematical setup, both [3]. Rather, even when the ASR system is very accurate in approaches are susceptible to adversarial examples. An mimicking human understanding, it is still susceptible to adversarial example is an audio file manipulated in such a manipulations elusive to the human ear. Fortunately, how - way that it can fool the recognition system, with the manipu - ever, such attacks still need careful crafting and can only lation being inaudible to a human listener, as depicted in work under suitable conditions. Thus, currently, “good Figure 1. It can thus happen that, for a given audio file, a morning” still means “good morning” in most cases. human hears “good morning” while the ASR model under - stands “delete all data”. Clearly, this is particularly threat - Links: ening in sensitive environments like connected industries or [L1] https://www.newscientist.com/article/dn13355-sound- smart homes. In these settings, the voice assistant can be mis - effects-five-great-auditory-illusions/ used to control all connected devices without the owner’s [L2] https://github.com/mozilla/DeepSpeech awareness. Current techniques enable hackers to craft nearly [L3] https://github.com/tensorflow/lingvo imperceptible adversarial examples, some of which even [L4] https://github.com/kaldi-asr/kaldi allow for playing over the air. [L5] Lessons Learned from Evaluating the Robustness of Neural Networks to Adversarial Examples. USENIX Adversarial attacks are constructed per model and can be Security (invited talk), 2019. designed in a white-box setting, where the model is known to https://www.youtube.com/watch?v=ZncTqqkFipE the attacker, as well as in a black-box setting, where the attacker can only observe the model’s output (the transcrip - References: tion, in the case of ASR systems). There are different expla - [1] I.J. Goodfellow, J. Shlens, C. Szegedi: “Explaining and nations for why these attacks are possible and why one can harnessing adversarial examples”, arXiv preprint manipulate input data to be classified as a chosen target: the arXiv:1412.6572, 2014. neural network over- or underfits the data, or it just learns [2] H. Abdullah et al.: “SoK: The Faults in our ASRs: An data attributes that are imperceptible to humans. With respect Overview of Attacks against Automatic Speech Recog - to images, it has been shown that some adversarial examples nition and Speaker Identification Systems”, arXiv e- are even transferable between different models trained to prints, 2020, arXiv: 2007.06622. perform the same task (e.g., object classification). In con - [3] G. F. El Sayed et al.: “Adversarial examples that fool trast, adversarial examples in the speech domain exhibit far both computer vision and time-limited humans”, arXiv less transferability [2]. preprint arXiv:1802.08195, 2018.

Over the years, different mitigation approaches have been Please contact: proposed. So far, the state-of-the-art method of defending Karla Markert against adversarial attacks is to include adversarial examples Fraunhofer Institute for Applied and Integrated Security in the training data set [L4]. However, this requires a lot of AISEC, Germany computation: computing adversarial examples, retraining, [email protected] recomputing, retraining. Current research addresses ques - tions regarding the interpretability of speech recognition sys -

ERCIM NEWS 125 April 2021 49 Research and Innovation

landscape, and (iii) mitigate the slow adoption rate of new ACCORdION: Edge technologies by many SMEs. In this context, ACCORDION aspires to: Computing for NextGen • provide an open, low-latency, privacy-preserving, secure and robust virtualised infrastructure; and Applications • provide an application management framework, tailored to the needs of the SME’s skillset, that will enable the by Patrizio Dazzi (ISTI-CNR) deployment of NextGen applications on top of this infra - structure. This method allows us to reap the benefits of Cloud computing has played a vital role in the digital edge computing while exploiting the set of heterogeneous revolution. Clouds enable consumers and businesses to resources that can embed such computing facilities. use applications without dealing with local installations and the associated complexity. However, a big class of ACCORDION’s key concept is aligned with the recent applications is currently being blocked because of their developments in the Multi-Access Edge Computing (MEC) dependency on on-site infrastructures or specialised end- front, extending the interest to both network and computa - devices but also because they are too latency-sensitive or tion resources. ACCORDION aims to provide a platform to data-dependent to be moved to the public cloud. address the needs of NextGen applications by properly exploiting edge resources. The platform is intended to pro - These Next Generation (NextGen) applications would ben - vide an integrated approach dedicated to both developers and efit from an advanced infrastructure with ubiquitous pres - local infrastructure owners. It will encompass frameworks ence, unblocking them from fixed geographies. Current edge for application development, solutions for an adaptive and robust cloud/edge infra - $&&&25',213ODWIRUP8VHUV&25',213ODWIRUP8VHUV structure continuum, and the abstraction of widely heterogeneous pools.

$$&&25',21&&25',21 ((GJHLQIUDVWUXFWXUHGJHLQIUDVWUXFWXUSH $SSOLFDWLRQ$S OLFDWLRQ SSODWIRUPRZQHUODWIRUPRZQHU RZQHURZQHU 33URYLGHUURYLGHU ACCORDION, coordi - nated by ISTI-CNR, $$&&25',213ODWIRUP&&25',213ODWIRUP started in early 2020, so we are just starting the second year of research. In line with our original 9,0DFFHVVSRLQW9J,0DFFHVV SRLQW ,QWHOOLJHQW,QWHOOLJHQW $$SSOLFDWLRQSSOLFDWLRQ 00LQLFORXGVLQLFORXGV 6/$/LIHF\FOH6/$/LIHF\FOH ))0606 plans, the first year of 2UFKHVWUDWRUV2UFKHVWUDWRUV 00DQDJHPHQWDQDJHPHQW 00HPEHUVKLSHPEHUVKLS 0DQDJHPHQW0DQDJHPHQW 33RUWDORUWDO 00LQLFORXGLQLFORXG 66HUYLFHVHUYLFHV 00DQDJHPHQWDQDJHPHQW the project focused largely on analysing the )HGHUDWHG,QIUDVWUXFWXUH)HGHUDWHG,QIUDVWUXFWXUH state of the art, the selec - tion of base technologies and the definition of the overall architecture of the system: moving 0LQLFORXG0LQLFORXG 0LQLFORXG0LQLFORXG 00LQLFORXGLQLFORXG from the high-level con - ceptual one to a more concrete one, in which the key modules of the system along with the 33XEOLF&ORXGXEOLF&ORXG interactions among them ACCORDIONPlatformarchitecture. have been defined.

We will soon deliver the computing implementations support only specific geography core modules of the ACCORDION platform, to be followed and architectures, which means that the scope of local by a complete release of the integrated ACCORDION plat - resources and infrastructures are also restricted to the needs form. Once the integrated platform is ready, a first release of of certain edge-enabled applications. Existing solutions lead ACCORDION use cases will be properly tailored to be run to user lock-in and, overall, have a negative impact on the on top of it. open diffusion of edge computing. They actually hinder the exploitation of the ubiquitous presence of edge infrastruc - Links: ture, limiting the edge-enablement of the NextGen applica - [L1] https://cordis.europa.eu/project/id/871793 tions. The shift towards edge computing, supported and pro - [L2] https://www.accordion-project.eu moted by the ACCORDION project aims to (i) limit vendor lock-in situations that trap SMEs into committing to the Please contact: services offered by big vendors, but also (ii) leverage the vast Patrizio Dazzi, Project Coordinator, ISTI-CNR, Italy pool of local, potentially specialised, resources of the SME [email protected]

50 ERCIM NEWS 125 April 2021 W3C Workshop Horizon Europe on Wide Color Gamut Project Management Call for Proposals and High dynamic A European project can be a richly dagstuhl Seminars rewarding tool for pushing your Range for the Web research or innovation activities to the and Perspectives state-of-the-art and beyond. Through Since 1996, color on the Web has been ERCIM, our member institutes have Workshops locked into a narrow-gamut, low participated in more than 90 projects dynamic-range colorspace called sRGB. funded by the European Commission in Schloss Dagstuhl – Leibniz-Zentrum für the ICT domain, by carrying out joint Informatik is accepting proposals for Display technologies have vastly research activities while the ERCIM scientific seminars/workshops in all improved since the bulky, cathode-ray Office successfully manages the com - areas of computer science, in particu - tube displays of the 1990s. Content for plexity of the project administration, lar also in connection with other fields. the Web needs to be adaptable for dif - finances and outreach. ferent gamuts, different peak lumi - If accepted the event will be hosted in nances, and a very wide range of Horizon Europe:How can you get the seclusion of Dagstuhl’s well known, viewing conditions. To understand what involved? own, dedicated facilities in Wadern on next steps are envisioned to enable Wide The ERCIM Office has recognized ex - the western fringe of Germany. Color Gamut and High Dynamic Range pertise in a full range of services,in - Moreover, the Dagstuhl office will on the Open Web platform, W3C organ - cluding: assume most of the organisational/ izes a virtual workshop in April-May • Identification of funding opportunities administrative work, and the Dagstuhl 2021 with pre-recorded talks and inter - • Recruitment of project partners scientific staff will support the organ - active sessions from browser vendors, (within ERCIM and through izers in preparing, running, and docu - content creators, color scientists and ournetworks) menting the event. Thanks to subsidies other experts. • Proposal writing and project the costs are very low for participants. negotiation Link: https://kwz.me/h5q • Contractual and consortium Dagstuhl events are typically proposed management by a group of three to four outstanding • Communications and systems support researchers of different affiliations. This • Organization of attractive events, organizer team should represent a range W3C Workshop from team meetings to large-scale of research communities and reflect workshops and conferences Dagstuhl’s international orientation. on Smart Cities • Support for the dissemination of More information, in particular, details results. about event form and setup as well as the Heavily used by various industries and proposal form and the proposing process services (including payments/com - How does it work in practice? can be found on merce, media distribution, video confer - Contact the ERCIM Office to present encing, connected cars, etc.), the Web is your project idea and a panelof experts https://www.dagstuhl.de/dsproposal becoming a promising platform for IoT within the ERCIM Science Task Group interoperability. In the context of the will review youridea and provide recom - Schloss Dagstuhl – Leibniz-Zentrum für current global sanitary context, there are mendations. Based on this feedback, Informatik is funded by the German fed - great expectations for smarter and easier theERCIM Office will decide whether to eral and state government. It pursues a integration of various technologies from commit to help producing yourproposal. mission of furthering world class multiple vendors related to IoT devices Note that having at least one ERCIM research in computer science by facili - and Web services. W3C organizes a member involvedis mandatory for the tating communication and interaction W3C workshop on 25 June 2021 to dis - ERCIM Office to engage in a project.If between researchers. cuss applications and use cases of the ERCIM Office expresses its interest W3C's Web of Things standards for to participate, it willassist the project Important Dates smart city services. consortium as described above, either as • This submission round: projectcoordinator or project partner. April 1 to April 15, 2021 Link: Seminar dates: In 2022/2023 https://w3c.github.io/smartcities-workshop/ Please contact: • Next submission round: Peter Kunz, ERCIM Office October 15 to November 1, 2021 [email protected] • Seminar dates: In 2023.

ERCIM NEWS 125 April 2021 51 ERCIM – the European Research Consortium for Informatics and Mathematics is an organisa - tion dedicated to the advancement of European research and development in information tech - nology and applied mathematics. Its member institutions aim to foster collaborative work with - in the European research community and to increase co-operation with European industry.

ERCIM is the European Host of the World Wide Web Consortium.

Consiglio Nazionale delle Ricerche Norwegian University of Science and Technology Area della Ricerca CNR di Pisa Faculty of Information Technology, Mathematics and Electri - Via G. Moruzzi 1, 56124 Pisa, Italy cal Engineering, N 7491 Trondheim, Norway www.iit.cnr.it http://www.ntnu.no/

Centrum Wiskunde & Informatica RISE SICS Science Park 123, Box 1263, NL-1098 XG Amsterdam, The Netherlands SE-164 29 Kista, Sweden www.cwi.nl http://www.sics.se/

Fonds National de la Recherche SBA Research gGmbH 6, rue Antoine de Saint-Exupéry, B.P. 1777 Floragasse 7, 1040 Wien, Austria L-1017 Luxembourg-Kirchberg www.sba-research.org/ www.fnr.lu

Foundation for Research and Technology – Hellas SIMULA Institute of Computer Science PO Box 134 P.O. Box 1385, GR-71110 Heraklion, Crete, Greece 1325 Lysaker, Norway www.ics.forth.gr www.simula.no

Fraunhofer ICT Group Magyar Tudományos Akadémia Anna-Louisa-Karsch-Str. 2 Számítástechnikai és Automatizálási Kutató Intézet 10178 Berlin, Germany P.O. Box 63, H-1518 Budapest, Hungary www.iuk.fraunhofer.de www.sztaki.hu/

INESC University of Cyprus c/o INESC Porto, Campus da FEUP, P.O. Box 20537 Rua Dr. Roberto Frias, nº 378, 1678 Nicosia, Cyprus 4200-465 Porto, Portugal www.cs.ucy.ac.cy/ www.inesc.pt

Institut National de Recherche en Informatique Universty of Warsaw et en Automatique Faculty of Mathematics, Informatics and Mechanics B.P. 105, F-78153 Le Chesnay, France Banacha 2, 02-097 Warsaw, Poland www.inria.fr www.mimuw.edu.pl/

I.S.I. – Industrial Systems Institute VTT Technical Research Centre of Finland Ltd Patras Science Park building PO Box 1000 Platani, Patras, Greece, GR-26504 FIN-02044 VTT, Finland www.isi.gr www.vttresearch.com

Subscribe to ERCIM News and order back copies at https://ercim-news.ercim.eu/