Zurich Open Repository and Archive University of Zurich Main Library Strickhofstrasse 39 CH-8057 Zurich www.zora.uzh.ch

Year: 2011

Frontiers in neuromorphic

Indiveri, G ; Horiuchi, T K

DOI: https://doi.org/10.3389/fnins.2011.00118

Posted at the Zurich Open Repository and Archive, University of Zurich ZORA URL: https://doi.org/10.5167/uzh-60627 Journal Article Published Version

The following work is licensed under a Publisher License.

Originally published at: Indiveri, G; Horiuchi, T K (2011). Frontiers in neuromorphic engineering. Frontiers in , 5:118. DOI: https://doi.org/10.3389/fnins.2011.00118 Specialty Grand Challenge Article published: 10 October 2011 doi: 10.3389/fnins.2011.00118 Frontiers in neuromorphic engineering

Giacomo Indiveri1* and Timothy K. Horiuchi 2

1 Institute of Neuroinformatics, University of Zurich and ETH Zurich, Zurich, Switzerland 2 Department of Electrical and , Institute for Systems , University of Maryland, College Park, MD, USA *Correspondence: [email protected]

Neurobiological processing systems are current characteristics of the transistor such as silicon retinas, visual motion sen- remarkable computational devices. They began in the mid 1980s with the collabo- sors, and silicon cochleas for a wide vari- use slow, stochastic, and inhomogeneous ration that sprung up between prominent ety of applications. In recent years, many computing elements and yet they outper- scientists Max Delbrück, , larger multi-chip neuromorphic systems form today’s most powerful computers at Carver Mead, and Richard Feynman (Hey, have begun to emerge that have raised new tasks such as vision, audition, and motor 1999). Inspired by graded synaptic trans- issues and challenges. These systems typically control, tasks that we perform nearly every mission in the retina, Mead sought to use comprise one or more neuromorphic sen- moment that we are awake without much the graded (analog) properties of transis- sors, interfaced to general-purpose neural conscious thought or concern. Despite the tors, rather than simply operating them as network chips using spiking silicon neurons vast amount of resources dedicated to the on–off (digital) switches. He showed that and dynamic synapses. research and development of computing, analog neuromorphic circuits share many The method used to transmit spikes information, and communication technol- common physical properties with protein across chip boundaries in these systems is ogies, today’s fastest and largest computers channels in neurons (Mead, 1989). As a con- based on the address-event representation are still not able to match biological sys- sequence, these types of circuits require far (AER; Mahowald, 1994). It is an asynchro- tems at robustly accomplishing real-world fewer transistors than digital approaches to nous digital communication protocol that tasks. While the specific algorithms and emulating neural systems. sends the address of the neuron that emitted representations that biological use Through the of Computation the event in real-time or close to real-time. are still largely unknown, it is clear that course at Caltech (led by Carver Mead, John The information being transmitted may instead of Boolean logic, precise digital Hopfield, and Richard Feynman), Mead be analog or digital, but must be commu- representations, and synchronous opera- (1989)’s textbook Analog VLSI and Neural nicated via spikes, thus raising the critical tions, nervous systems use hybrid analog/ Systems, and the creation of the Telluride and exciting issue of signal encoding that digital components, distributed represen- Neuromorphic Engineering Workshop, the is currently a very active topic in neurosci- tations, massively parallel mechanisms, field of Neuromorphic Engineering was ence. Signals can be encoded in the mean combine communications with established. Prominent in the early expansion frequency (rate) of spikes, in their precise and computation, and make extensive use of the field were scientists and engineers such timing with respect to a time reference, or of adaptation, self-organization, and learn- as Christof Koch, Terry Sejnowski, Rodney in the population response. In fact, multiple ing. On the other hand, as with many suc- Douglas, Andreas Andreou, Paul Mueller, Jan signals can be simultaneously encoded in a cessful man-made systems, it is clear that van der Spiegel, and Eric Vittoz, training a single spike train. Once on a digital bus, the biological brains have been co-designed generation of cross-disciplinary students. address-events can be remapped to multiple with the body to operate under a specific It has been argued that neuromorphic destinations using commercially available range of conditions and assumptions about circuits are ideal for developing a new gen- synchronous or custom asynchronous pro- the world. eration of computing technologies that cessing. Digital AER infrastructures allow us Understanding the computational prin- use the same organizing principles of the to construct large multi-chip networks with ciples used by the and how they are biological nervous system (Douglas et al., nearly arbitrary connectivity and to dynami- physically embodied is crucial for develop- 1995; Boahen, 2005; Sarpeshkar, 2006). In cally reconfigure the network topology for ing novel computing paradigms and guid- addition to the computations of a single experimentation. By using analog circuits ing a new generation of technologies that neuron, many neuromorphic circuits also for local computations on-chip and digi- can combine the strengths of industrial- utilize spiking representations for commu- tal circuits for long-distance communica- scale electronics with the computational nication, learning and memory, and com- tion (off-chip), neuromorphic systems can performance of brains. putation. The use of asynchronous spike- or exploit the best of both worlds. digital event-based representations in elec- Another distinguishing feature of neuro- Neuromorphic Engineering tronic systems can be energy-efficient and morphic engineering has been the integra- While the history of implementing electronic fault-tolerant, making them ideal for build- tion of fine-grained synaptic modification models of neural circuits extends back to the ing modular systems and creating complex mechanisms that both enable these net- construction of perceptrons (Rosenblatt, hierarchies of computation. works to change their behavior with experi- 1958) and retinas (Fukushima et al., 1970), The most successful neuromorphic sys- ence (as is ubiquitous in biological nervous the modern wave of research utilizing VLSI tems to date have been single chip devices systems) and to implicitly overcome the technology and ­emphasizing the non-­linear that emulate peripheral sensory transduction­ inherent device parameter variability

www.frontiersin.org October 2011 | Volume 5 | Article 118 | 1 Indiveri and Horiuchi Neuromorphic engineering

found in all manufacturing technologies While the majority of cognitive archi- • neural computation, involving studies of whether silicon or mechanical. The most tectures and their software implemen- spiking winner-take-all networks, attrac- prominent storage mechanisms have been: tations have avoided detailed neural tor networks, mean-field , spike- on-chip capacitance, on-chip floating-gate implementations, due to limited computa- based learning mechanisms, probabilistic charge storage, and off-chip AER remap- tional power and the assumption that the graphical models, cortical development, ping of the network to either dynamically details of single spiking neurons are not and self-constructing principles; change the connectivity or to implement important at this level, a growing number • biologically plausible cognitive architec- stochastic spike delivery. To implement bio- of research groups worldwide have begun tures for studying attention, working logically plausible learning rules (e.g., spike- to consider the consequences of biologi- memory, state-dependent computa- timing dependent plasticity), many of these cally plausible implementations at both tion, action selection mechanisms, implementations also incorporate learning the level of neural fields and single spiking planning, and multi-agent interaction. circuits directly at the synapse. neurons. By providing real-time spiking implementations of core neural circuits, Through this journal, we intend to Frontiers in Neuromorphic Engineering neuromorphic engineering will play an encourage the presentation of these diverse At its heart, neuromorphic engineering is important role in the development and perspectives, technical approaches, and about the real-time interaction of the algo- fielding of biologically relevant working goals, to facilitate the development of neu- rithm with its physical implementation and models of cognition interacting with the romorphic cognitive systems, and reach the environment in solving tasks. This syn- real-world. new frontiers in neuromorphic engineering. ergy is easy to appreciate at the sensory and One of the Grand Challenges of motor interfaces with the world, but more Neuromorphic Engineering is to References Boahen, K. A. (2005). Neuromorphic microchips. Sci. subtle and interesting when considering ­demonstrate cognitive systems using Am. 292, 56–63. cognitive-level tasks. hardware neural processing architectures Douglas, R. J., Mahowald, M. A., and Mead, C. (1995). With increasing knowledge of what integrated with physical bodies (e.g., Neuromorphic analogue VLSI. Annu. Rev. Neurosci. single neurons and their synapses can do humanoid robots) that can solve every- 18, 255–281. computationally, the desire for more sophis- day tasks in real-time. To be successful Fukushima, K., Yamaguchi, Y., Yasuda, M., and Nagata, S. (1970). An electronic model of the retina. Proc. IEEE ticated implementation technologies has in this ambitious endeavor, an integrated 58, 1950–1951. grown. At present, new technologies such multi-disciplinary approach is critical that Hey, T. (1999). Richard Feynman and computation. as nano-scale transistors, quantum devices, brings together research in: Contemp. Phy. 40, 257–265. organic electronics, memristors, phase- Mahowald, M. (1994). An Analog VLSI System for change materials, 3D integrated circuits, • VLSI circuits and systems for imple- Stereoscopic Vision. Boston, MA: Kluwer. Mead, C. A. (1989). Analog VLSI and Neural Systems. and electro-active polymers for actuation menting hardware models of neural Reading, MA: Addison-Wesley. are all promising directions for research. processing systems, mixed analog/digi- Rosenblatt, F. (1958). The perceptron: a probabilistic Neuromorphic engineering now aims tal asynchronous AER communication model for information storage and organization in to use these technologies for develop- infrastructures, spike-based sensory– the brain. Psychol. Rev. 65, 386–408. ing larger-scale neural processing systems motor systems, and event-driven pro- Sarpeshkar, R. (2006). Brain power – borrowing from makes for low power computing – bionic ear. and move from the predominantly feed-­ cessing methods; IEEE Spectr. 43, 24–29. forward, reactive neuromorphic systems of • emerging technologies including 3D the past to adaptive behaving ones that can VLSI, nanotechnologies, phase-change Received: 19 July 2011; accepted: 14 September 2011; pub- be considered cognitive. For example, a key materials, and memristive devices, lished online: 10 October 2011. mechanism in cognition, selective attention, applied to the construction of low- Citation: Indiveri G and Horiuchi TK (2011) Frontiers in neuromorphic engineering. Front. Neurosci. 5:118. doi: has long been part of the neuromorphic power neuromorphic systems; 10.3389/fnins.2011.00118 engineering toolkit, but has largely operated • robotic platforms and control with parti- This article was submitted to Frontiers in Neuromorphic as a bottom-up process, operating on short- cular focus on new actuators and mate- Engineering, a specialty of Frontiers in Neuroscience. term information and memory. Expanding rials, compliant systems, contraction Copyright © 2011 Indiveri and Horiuchi. This is an open- its role in top-down behavior (e.g., guiding theory and controllability of complex access article subject to a non-exclusive license between the authors and Frontiers Media SA, which permits use, dis- the learning of more abstract concepts) will systems, and on the computational role tribution and reproduction in other forums, provided the be important for understanding and imple- of the physical body in locomotion and original authors and source are credited and other Frontiers menting context-dependent behavior. active sensing; conditions are complied with.

Frontiers in Neuroscience | Neuromorphic Engineering October 2011 | Volume 5 | Article 118 | 2