DOCSLIB.ORG

Arxiv:2108.12596V1 [Cs.LG] 28 Aug 2021

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Arxiv:2108.12596V1 [Cs.LG] 28 Aug 2021 Representation Memorization for Fast Learning New Knowledge without Forgetting Fei Mi1∗, Tao Lin2, , Boi Faltings2 1Huawei Noah’s Ark Lab, 2Swiss Federal Institute of Technology in Lausanne (EPFL) [email protected], ftao.lin,boi.faltingsg@@epfl.ch Abstract limited, which prevents neural networks from being trained to achieve reasonable accuracy Wang et al. [2020]. The ability to quickly learn new knowledge (e.g. The catastrophic forgetting challenge is recently studied new classes or data distributions) is a big step in the context of “continually/incrementally” learning a se- towards human-level intelligence. In this paper, quence of tasks. Various regularization-base methods Kirk- we consider scenarios that require learning new patrick et al. [2017]; Rebuffi et al. [2017]; Castro et al. [2018] classes or data distributions quickly and incremen- and memory-based approaches Grave et al. [2017b]; Merity tally over time, as it often occurs in real-world dy- et al. [2017]; Rebuffi et al. [2017] have been proposed to mit- namic environments. We propose “Memory-based igate catastrophic forgetting. The sample efficiency challenge Hebbian Parameter Adaptation” (Hebb) to tackle is recently studied in the context of few-shot learning; a pop- the two major challenges (i.e., catastrophic forget- ular approach is “meta-learning” Finn et al. [2017] that learns ting and sample efficiency) towards this goal in a over a bunch of specifically structured meta-tasks. unified framework. To mitigate catastrophic for- However, existing methods often tackle these two chal- getting, Hebb augments a regular neural classifier lenges separately. To this end, we propose a method called with a continuously updated memory module to Memory-based Hebbian Parameter Adaptation (Hebb) to store representations of previous data. To improve tackle them in a unified framework. Hebb makes use of a sample efficiency, we propose a parameter adapta- memory component similar to Merity et al. [2017]; Sprech- tion method based on the well-known Hebbian the- mann et al. [2018] that stores previous input representations ory Hebb [1949], which directly “wires” the output to mitigate catastrophic forgetting. To improve sample ef- network’s parameters with similar representations ficiency on new knowledge, we propose a parameter adap- retrieved from the memory. We empirically verify tation procedure based on the well-known Hebbian theory the superior performance of Hebb through exten- Hebb [1949] during inference. It directly wires similar rep- sive experiments on a wide range of learning tasks resentations retrieved from the memory to the corresponding (image classification, language model) and learn- parameters of the classifier’s output network. The memory ing scenarios (continual, incremental, online). We accessing operation and the parameter adaptation procedure demonstrate that Hebb effectively mitigates catas- can be easily computed such that Hebb can be easily plugged trophic forgetting, and it indeed learns new knowl- into different neural classifiers and learning scenarios. edge better and faster than the current state-of-the- Besides the standard continual learning setting Rebuffi et art. al. [2017] to evaluate the ability to mitigate catastrophic for- getting, two other learning scenarios are considered to evalu- arXiv:2108.12596v2 [cs.LG] 10 Sep 2021 ate the ability to deal with the sample efficiency challenge. In 1 Introduction the first incremental learning scenario, a pre-trained classifier In real-life machine learning applications, new knowledge is initially trained on a small dataset containing new knowl- (e.g., new classes or data distributions) arrive gradually over edge. Then it is fixed to be evaluated w.r.t. future observa- time. The ability to quickly learn and accumulate new knowl- tions. In the second online adaptation scenario, we have no edge without forgetting old ones is a hallmark of artificial in- data on new knowledge initially, and the pre-trained classifier telligence. Two major challenges prevent standard neural net- needs to continuously learn new knowledge through a single works to be applied towards this goal. (1) catastrophic for- pass over new data in an online manner. These two learning getting: continually incorporating new knowledge requires scenarios are both practically critical. For example, the incre- additional training and often reduces the performance on old mental learning scenario could simulate that a robot is shown ones learned before McCloskey and Cohen [1989]; (2) sam- some images of new objects before they appear in its routine ple efficiency: the amount of data on new knowledge is often tasks. In the online adaptation scenarios, a robot has to deal with new objects continuously. ∗This work was mainly done when Fei Mi was a Ph.D. in EPFL Through extensive experiments on a wide range of learn- ing tasks (image classification, language model) and learn- to bind representative patterns to their corresponding con- ing scenarios (continual, incremental, online), we empirically cepts or labels. demonstrate: (i) Hebb can be easily plugged into different Recent approaches apply Hebbian theory to every single neural classifiers and learning scenarios with trivial compu- neuron connection for fast network weight learning. Ba et al. tation overhead. (ii) Hebb effectively mitigates catastrophic [2016] proposes a fast weight to augment the standard com- forgetting, indicated by its superior performance compared putation of RNNs. The fast weight is defined as the running to MbPA Sprechmann et al. [2018] and EWC Kirkpatrick et average of the outer product of two hidden states in RNNs. It al. [2017] in the continual learning setting. (iii) More im- is multiplied to the current state and it is continuously updated portantly, Hebb notably improves sample efficiency for fast to allow each new hidden state to be attracted to recent hidden learning new classes and data distributions. It outperforms states. Miconi et al. [2018] later augments the traditional con- various state-of-the-art methods in both incremental learning nection between two neurons in general neural networks with and online adaptation scenarios, especially on new or infre- a Hebbian trace. The Hebbian trace between two neurons quent classes. is defined as a running average of the scalar product of the first neuron’s activation in the last timestamp and the second 2 Related Work neuron’s activation in the current timestamp. The Hebbian 2.1 Memory-augmented Neural Networks trace is merged with the standard neuron connection through a differentiable plasticity optimized by SGD. The later exten- Recently, various memory modules (M) have been proposed sion Miconi et al. [2019] introduces a term parametrized by to augment neural networks for remembering long-term in- another neural network to learn how fast should new informa- formation Graves et al. [2014]; Grave et al. [2017b,a]; Mer- tion be incorporated. ity et al. [2017] or learning infrequent patterns Santoro et al. Instead of applying the Hebbian theory to every single neu- [2016]; Kaiser et al. [2017]; Sprechmann et al. [2018]; Mi ron connection, we use it to directly wire the activation input and Faltings [2020]. to the output layer with the corresponding class label for fast M There are many variants of how to read from and mix binding new classes. Similar perspectives are recently pro- M the entries retrieved from with the network computa- posed. For example, Munkhdalai and Trischler [2018] aug- tion. One approach is through some differentiable context- ments the layer preceding the Softmax layer with the Hebbian et al. et al. based lookup mechanisms Graves [2014]; Santoro updates followed by a nonlinear activation for meta-learning. [2016] for learning to match the current activation to past ac- Rae et al. [2018] proposes a Hebbian Softmax layer during M tivations stored in . However, these mechanisms often re- the normal model training phase to better learn infrequent vo- M quire strong memory supervision, and the size of the has to cabularies in language modeling tasks. The Hebbian update be fixed. Another approach is using a simple mixture model. rule proposed in our paper is motivated by Rae et al. [2018], In this case, a non-parametric prediction is computed based yet our Hebbian update rule is only applied to relevant entries M on the similarity between the entries in and the current in- retrieved from a continuously updated memory module for put. The neural network’s prediction is directly interpolated the purpose of fast learning new knowledge in an incremental M with the non-parametric prediction from . This approach or online manner. has been shown simple but effective for language model- ing Grave et al. [2017b,a], neural machine translation Tu et al. [2018], image classification Orhan [2018], and recommen- dation Mi and Faltings [2020]. Recently, Sprechmann et al. 3 Memory-based Hebbian Parameter [2018] introduces MbPA to use nearest neighbors retrieved Adaptation from M for parameter adaptation during model inference for the fast acquisition of new knowledge. MbPA++ de Mas- This section introduces the Memory-based Hebbian Param- son d’Autume et al. [2019] improves MbPA to better mitigate eter Adaptation (Hebb) method to help standard neural clas- catastrophic forgetting through better memory management sifiers mitigate catastrophic forgetting and improve sample during training. The framework proposed in this paper is mo- efficiency. First, we introduce a memory component to store tivated by Sprechmann et al. [2018], and it mainly improves representations of past data. Then, we introduce the Hebbian MbPA for better learning new knowledge. update for fast learning new knowledge during inference, and we compare it with state-of-the-art (MbPA Sprechmann et al. 2.2 Hebbian Learning [2018]). Lastly, we introduce a dynamic interpolation of the Hebbian theory Hebb [1949] is a neuroscientific theory at- proposed Hebbian update and MbPA. tempting to explain “synaptic plasticity”, i.e., the adaptation of brain neurons during the learning process. For artificial Background Neural classifiers can be visualized by two neural networks, Hebbian theory describes a method of de- parts.
Load more

Recommended publications
  • Song Exposure Affects HVC Ultrastructure in Juvenile Zebra Finches
    Song Exposure affects HVC Ultrastructure in Juvenile Zebra Finches Houda G Khaled Advisor: Dr. Sharon M.H. Gobes Department of Neuroscience Submitted in Partial Fulfillment of the Prerequisite for Honors in Biochemistry May 2016 © 2016 Houda Khaled Table of Contents ABSTRACT ........................................................................................................................................................ 2 ACKNOWLEDGEMENTS .............................................................................................................................. 3 INTRODUCTION ............................................................................................................................................. 4 I. SYNAPTIC PLASTICITY IN LEARNING AND MEMORY ................................................................................ 4 Basic principles of synapse structure ............................................................................................................. 4 Changes in synapse number, size and shape ................................................................................................ 6 II. SONG ACQUISITION AS A LEARNING PARADIGM .................................................................................. 7 The song system ................................................................................................................................................. 9 Structural synaptic plasticity in the song system .....................................................................................
    [Show full text]
  • Interplay of Multiple Plasticities and Activity Dependent Rules: Data, Models and Possible Impact on Learning
    Interplay of multiple plasticities and activity dependent rules: data, models and possible impact on learning Gaetan Vignoud, Alexandre Mendes, Sylvie Perez, Jonathan Touboul*, Laurent Venance* (* equal contributions, email: fi[email protected]) College` de France, Center for Interdisciplinary Research in Biology, 11 Place Marcelin Berthelot, 75005 Paris, France Abstract the millisecond timing of the paired activities on either side of Synaptic plasticity, the activity-dependent evolution neuronal the synapse (Feldman, 2012). As such, plasticity in the stria- connectivity, is admitted to underpin learning and memory. A tum has been widely studied and a variety of spike-timing de- paradigmatic plasticity depending on the spike timings of cells pendent plasticity were observed in response to a large num- on both sides of the synapse (STDP), was experimentally ev- ber of presentations of a fixed pre- and post-synaptic stim- idenced through multiple repetitions of fixed pre- and post- ulation pattern. Those plasticities were attributed to a di- synaptic spike patterns. Theoretical and experimental com- verse neurotransmitters and pathways including for instance munities often implicitly admit that plasticity is gradually estab- endocannabinoids (eCB), NMDA receptors, AMPA receptors lished as stimulus patterns are presented. Here, we evaluate or voltage-sensitive calcium channels. Moreover, the vast ma- this hypothesis experimentally and theoretically in the stria- jority of the studies focused on cortical inputs, and much
    [Show full text]
  • A Theory of Consumer Fraud in Market Economies
    A THEORY OF FRAUD IN MARKET ECONOMIES ADISSERTATIONIN Economics and Social Science Presented to the Faculty of the University of Missouri-Kansas City in partial fulfillment of the requirements for the degree DOCTOR OF PHILOSOPHY by Nicola R. Matthews B.F.A, School of Visual Arts, 2002 M.A., State University College Bu↵alo, 2009 Kansas City, Missouri 2018 ©2018 Nicola R. Matthews All Rights Reserved A THEORY OF FRAUD IN MARKET ECONOMIES Nicola R. Matthews, Candidate for the Doctor of Philosophy Degree University of Missouri-Kansas City, 2018 ABSTRACT Of the many forms of human behavior, perhaps none more than those of force and fraud have caused so much harm in both social and ecological environments. In spite of this, neither shows signs of weakening given their current level of toleration. Nonetheless, what can be said in respect to the use of force in the West is that it has lost most of its competitiveness. While competitive force ruled in the preceding epoch, this category of violence has now been reduced to a relatively negligible degree. On the other hand, the same cannot be said of fraud. In fact, it appears that it has moved in the other direction and become more prevalent. The causes for this movement will be the fundamental question directing the inquiry. In the process, this dissertation will trace historical events and methods of control ranging from the use of direct force, to the use of ceremonies and rituals and to the use of the methods of law. An additional analysis of transformation will also be undertaken with regards to risk-sharing-agrarian-based societies to individual-risk-factory-based ones.
    [Show full text]
  • Arxiv:1711.09601V4 [Cs.CV] 5 Oct 2018 Unlabeled Data Towards What the Network Needs (Not) to Forget, Which May Vary Depending on Test Conditions
    Memory Aware Synapses: Learning what (not) to forget Rahaf Aljundi1, Francesca Babiloni1, Mohamed Elhoseiny2, Marcus Rohrbach2, and Tinne Tuytelaars1 1 KU Leuven, ESAT-PSI, iMEC, Belgium 2 Facebook AI Research Abstract. Humans can learn in a continuous manner. Old rarely utilized knowl- edge can be overwritten by new incoming information while important, frequently used knowledge is prevented from being erased. In artificial learning systems, lifelong learning so far has focused mainly on accumulating knowledge over tasks and overcoming catastrophic forgetting. In this paper, we argue that, given the limited model capacity and the unlimited new information to be learned, knowl- edge has to be preserved or erased selectively. Inspired by neuroplasticity, we propose a novel approach for lifelong learning, coined Memory Aware Synapses (MAS). It computes the importance of the parameters of a neural network in an unsupervised and online manner. Given a new sample which is fed to the net- work, MAS accumulates an importance measure for each parameter of the net- work, based on how sensitive the predicted output function is to a change in this parameter. When learning a new task, changes to important parameters can then be penalized, effectively preventing important knowledge related to previous tasks from being overwritten. Further, we show an interesting connection between a local version of our method and Hebb’s rule, which is a model for the learning process in the brain. We test our method on a sequence of object recognition tasks and on the challenging problem of learning an embedding for predicting <subject, predicate, object> triplets. We show state-of-the-art performance and, for the first time, the ability to adapt the importance of the parameters based on arXiv:1711.09601v4 [cs.CV] 5 Oct 2018 unlabeled data towards what the network needs (not) to forget, which may vary depending on test conditions.
    [Show full text]
  • Building a “Cross-Roads Discipline” at Mcgill University: a History of Early Experimental Psychology in Postwar Canada
    BUILDING A “CROSS-ROADS DISCIPLINE” AT MCGILL UNIVERSITY: A HISTORY OF EARLY EXPERIMENTAL PSYCHOLOGY IN POSTWAR CANADA ERIC OOSENBRUG A dissertation submitted to the Faculty of Graduate Studies in partial fulfillment of the requirements for the degree of Doctor of Philosophy Graduate Program in Psychology. Graduate Program in Psychology York University Toronto, Ontario October 2020 © Eric Oosenbrug, 2020 Abstract This dissertation presents an account of the development of psychology at McGill University from the late nineteenth century through to the early 1960s. The department of psychology at McGill represents an alternative to the traditional American-centered narrative of the cognitive revolution and later emergence of the neurosciences. In the years following World War II, a series of psychological experiments established McGill as among the foremost departments of psychology in North America. This thesis is an institutional history that reconstructs the origins, evolution, and dramatic rise of McGill as a major center for psychological research. The experiments conducted in the early 1950s, in the areas of sensory restriction, motivation, and pain psychology, were transformative in their scope and reach. Central to this story is Donald O. Hebb, author of The Organization of Behavior (1949), who arrived at McGill in 1947 to find the charred remains of a department. I argue that the kind of psychology Hebb established at McGill was different from most departments in North America; this is developed through a number of interwoven storylines focused on the understanding of a particular character of McGill psychology - a distinctive “psychological style” - and its broader historical importance for Canadian psychology, for North American psychology, and for psychology across the globe.
    [Show full text]
  • Neuromodulated Dopamine Plastic Networks for Heterogeneous Transfer Learning with Hebbian Principle
    S S symmetry Article Neuromodulated Dopamine Plastic Networks for Heterogeneous Transfer Learning with Hebbian Principle Arjun Magotra and Juntae Kim * Department of Computer Science and Engineering, Dongguk University, 30, Pildong-ro 1-gil, Jung-gu, Seoul 04620, Korea; [email protected] * Correspondence: [email protected] Abstract: The plastic modifications in synaptic connectivity is primarily from changes triggered by neuromodulated dopamine signals. These activities are controlled by neuromodulation, which is itself under the control of the brain. The subjective brain’s self-modifying abilities play an essential role in learning and adaptation. The artificial neural networks with neuromodulated plasticity are used to implement transfer learning in the image classification domain. In particular, this has application in image detection, image segmentation, and transfer of learning parameters with significant results. This paper proposes a novel approach to enhance transfer learning accuracy in a heterogeneous source and target, using the neuromodulation of the Hebbian learning principle, called NDHTL (Neuromodulated Dopamine Hebbian Transfer Learning). Neuromodulation of plasticity offers a powerful new technique with applications in training neural networks implementing asymmetric backpropagation using Hebbian principles in transfer learning motivated CNNs (Convolutional neural networks). Biologically motivated concomitant learning, where connected brain cells activate positively, enhances the synaptic connection strength between the network neurons. Using the NDHTL algorithm, the percentage of change of the plasticity between the neurons of the CNN layer is directly managed by the dopamine signal’s value. The discriminative nature of transfer learning Citation: Magotra, A.; Kim, J. fits well with the technique. The learned model’s connection weights must adapt to unseen target Neuromodulated Dopamine Plastic datasets with the least cost and effort in transfer learning.
    [Show full text]
  • Machine Awareness
    Wright State University CORE Scholar International Symposium on Aviation International Symposium on Aviation Psychology - 2019 Psychology 5-7-2019 Machine Awareness Steven D. Harbour Jeffery D. Clark William D. Mitchell Krishnamurthy V. Vemuru Follow this and additional works at: https://corescholar.libraries.wright.edu/isap_2019 Part of the Other Psychiatry and Psychology Commons Repository Citation Harbour, S. D., Clark, J. D., Mitchell, W. D., & Vemuru, K. V. (2019). Machine Awareness. 20th International Symposium on Aviation Psychology, 480-485. https://corescholar.libraries.wright.edu/isap_2019/81 This Article is brought to you for free and open access by the International Symposium on Aviation Psychology at CORE Scholar. It has been accepted for inclusion in International Symposium on Aviation Psychology - 2019 by an authorized administrator of CORE Scholar. For more information, please contact [email protected]. MACHINE AWARENESS Dr. Steven D. Harbour Dr. Jeffery D. Clark Mr. William D. Mitchell Riverside Research 2640 Hibiscus Way, Beavercreek, Ohio 45431, USA Dr. Krishnamurthy V. Vemuru Riverside Research 2900 Crystal Dr., Arlington, Virginia 22202, USA Current and future research that embodies a pathway to achieving machine common sense (MCS), including Capsule Neural Networks, Hebbian Plasticity Theory, Dual Process Theory, and machine awareness (MA). The final frontier may well involve a framework that is capable of machine curiosity, exploration, automatic self- direction and adaptation. The artificial intelligence (AI) system of the future will possess an innate curiosity and explore its own environment to gain knowledge, exhibiting a basic element of human cognition and awareness. The resulting MA system will possess inherent self-driven curiosity and related entropy in the decision space as it explores the environment in much the same manner as humans do.
    [Show full text]
  • Evidence for a Causal Inverse Model in an Avian Cortico-Basal Ganglia Circuit
    Evidence for a causal inverse model in an avian cortico-basal ganglia circuit Nicolas Gireta,b, Joergen Kornfelda, Surya Gangulic, and Richard H. R. Hahnlosera,b,1 aInstitute of Neuroinformatics and bNeuroscience Center Zurich, University of Zurich and ETH Zurich, 8057 Zurich, Switzerland; and cDepartment of Applied Physics, Stanford University, Stanford, CA 94305-5447 Edited by Thomas C. Südhof, Stanford University School of Medicine, Stanford, CA, and approved March 7, 2014 (received for review September 11, 2013) Learning by imitation is fundamental to both communication and nucleus of the anterior nidopallium (LMAN) forms the output of social behavior and requires the conversion of complex, nonlinear a basal-ganglia pathway involved in generating subtle song vari- sensory codes for perception into similarly complex motor codes ability (14–17). HVC neurons produce highly stereotyped firing for generating action. To understand the neural substrates un- patterns during singing (18, 19), whereas LMAN neurons produce derlying this conversion, we study sensorimotor transformations highly variable patterns (15, 20, 21). in songbird cortical output neurons of a basal-ganglia pathway To gain insights into LMAN’s role in motor control, we consider involved in song learning. Despite the complexity of sensory and recent theoretical work that establishes a conceptual link between motor codes, we find a simple, temporally specific, causal corre- inverse models and vocal-auditory mirror neurons (22–24). We spondence between them. Sensory neural responses to song consider three (nonexhaustive) possibilities about the flow of playback mirror motor-related activity recorded during singing, sensory information into motor areas. First, auditory afferents with with a temporal offset of roughly 40 ms, in agreement with short some feature sensitivity could map onto motor neurons involved in feedback loop delays estimated using electrical and auditory generating those same features (Fig.
    [Show full text]
  • Term Memory by ©2017 Liying Li
    Drosophila CPEB, Orb2, a Putative Biochemical Engram of Long- term Memory By ©2017 Liying Li Submitted to the graduate degree program in Molecular and Integrative Physiology and the Graduate Faculty of the University of Kansas in partial fulfillment of the requirements for the degree of Doctor of Philosophy __________________________ Co-Chair: Dr. Kausik Si __________________________ Co-Chair: Dr. Paul Cheney __________________________ Dr. Tatjana Piotrowski __________________________ Dr. Hiroshi Nishimune __________________________ Dr. Matthew Gibson Date Defended: June 22nd, 2017 The Dissertation Committee for Liying Li certifies that this is the approved version of the following dissertation: Drosophila CPEB, Orb2, a Putative Biochemical Engram of Long- term Memory _______________________________ Co-Chair: Dr. Kausik Si _______________________________ Co-Chair: Dr. Paul Cheney Date Approved: July 18th, 2017 ii Abstract How a transient experience creates an enduring yet dynamic memory remains a fundamental unresolved issue in studies of memory. Experience-dependent aggregation of the RNA-binding protein CPEB/Orb2 is one of the candidate mechanisms of memory maintenance. Here, using tools that allow rapid and reversible inactivation of Orb2 protein I find that Orb2 activity is required for encoding and recall of memory. Blocking the Orb2 oligomerization process by interfering with the protein phosphorylation pathway or expressing an anti-amyloidogenic peptide impairs long-term memory. Facilitating Orb2 aggregation by a DNA-J family chaperone, JJJ2, enhances the animal’s capacity to form long-term memory. Finally, I have developed tools to visualize training-dependent aggregation of Orb2. I find that aggregated Orb2 in subset of mushroom body neurons can serve as a “molecular signature” of memory and predict memory strength.
    [Show full text]
  • Fodor and Aquinas: the Architecture of the Mind and the Nature of Concept Acquisition
    FODOR AND AQUINAS: THE ARCHITECTURE OF THE MIND AND THE NATURE OF CONCEPT ACQUISITION A Dissertation submitted to the Faculty of the Graduate School of Arts and Sciences of Georgetown University in partial fulfillment for the requirements for the degree of Doctor of Philosophy in Philosophy By Justyna M. Japola, M.A. Georgetown University, Washington DC December 2009 Copyright 2009 by Justyna M. Japola All Rights Reserved ii FODOR AND AQUINAS: THE ARCHITECTURE OF THE MIND AND THE NATURE OF CONCEPT ACQUISITION Justyna M. Japola, M.A. Thesis Advisor: Wayne A. Davis, Ph. D. ABSTRACT Fodor (1975 and 1981b) explains the paradigm empiricist method of concept acquisition as consisting in forming and testing hypotheses about objects that fall under a concept. This method, he notices, can only work for complex concepts, because we have to possess some concepts in order to form hypotheses. If so, then none of our simple (or primitive) concepts can be learned. If we still have them then they must be innate. Aquinas, on the other hand, is famous for his opposition to Platonic nativism, and is universally considered an empiricist with respect to cognition. In my dissertation I show that Fodor’s and Aquinas’s accounts of the architecture of the mind are quite similar. I argue that because one's position in the empiricism-nativism debate should be a function of one's account of the architecture of the mind, Fodor and Aquinas should be on the same side of the debate. My claim is that they should be on the side of nativism, but not the kind of radical concept nativism that Fodor is famous for.
    [Show full text]
  • Hopfield Networks
    Hopfield Networks Jessica Yarnall John Hopfield • Son of two physicists • Earned PhD in physics from Cornell University in 1958 • Currently a professor of molecular biology at Princeton University • Developed a model in 1982 to explain how memories are recalled by the brain Now known as the Hopfield model Hebbian Theory and Associative Memory Hebbian Theory was introduced in 1949 to explain associative learning “Neurons that fire together, wire together.” Associative memory is the ability to learn and remember the relationship between two unrelated things Hopfield Network • Network is trained to store a number of patterns or memories • Can recognize partial or corrupted information about a pattern and returns the closest pattern or best guess • Comparable to human brain, it has stability in pattern recognition Hopfield Network • Single-layered, recurrent network Neurons are fully connected • Symmetric weights Given neurons i and j, wij = wji • Neuron is either on (firing) or off (not firing) On corresponds to 1 Off corresponds to -1 Training • Hebbian learning rule: Incremental – does not require information about patterns previously learnt Local - uses information from both nodes whose connectivity weight is updated • Formula: • Where µ εx is the state of node x in pattern µ Updating Hopfield Networks • Generally update nodes asynchronously Node chosen randomly • Update Rule: • Example: • http://faculty.etsu.edu/knisleyj/neural/ Energy Function • The energy function either decreases or stays the same with asynchronous updating • Energy Function: • Converges to a local minima Problems with Hopfield Networks • The more complex the thing being recalled, the more pixels and weights you’ll need • Spurious states References • https://www.fi.edu/laureates/john-j-hopfield • https://humans101.wordpress.com/tag/associative-memory/ • https://www.doc.ic.ac.uk/~ae/papers/Hopfield-networks-15.pdf • https://www.doc.ic.ac.uk/~sd4215/hopfield.html • https://www.youtube.com/watch?v=gfPUWwBkXZY.
    [Show full text]
  • Researching the Lived Experience of Individuals Practicing Self-Directed Neuroplasticity
    St. Catherine University SOPHIA Master of Arts in Holistic Health Studies Research Papers Holistic Health Studies 5-2019 Changing Brains, Changing Lives: Researching the Lived Experience of Individuals Practicing Self-Directed Neuroplasticity Tim Klein St. Catherine University, [email protected] Beth Kendall St. Catherine University, [email protected] Theresa Tougas St. Catherine University, [email protected] Follow this and additional works at: https://sophia.stkate.edu/ma_hhs Part of the Alternative and Complementary Medicine Commons Recommended Citation Klein, Tim; Kendall, Beth; and Tougas, Theresa. (2019). Changing Brains, Changing Lives: Researching the Lived Experience of Individuals Practicing Self-Directed Neuroplasticity. Retrieved from Sophia, the St. Catherine University repository website: https://sophia.stkate.edu/ma_hhs/20 This Thesis is brought to you for free and open access by the Holistic Health Studies at SOPHIA. It has been accepted for inclusion in Master of Arts in Holistic Health Studies Research Papers by an authorized administrator of SOPHIA. For more information, please contact [email protected]. Running head: LIVED EXPERIENCE OF SELF-DIRECTED NEUROPLASTICITY Changing Brains, Changing Lives: Researching the Lived Experience of Individuals Practicing Self-Directed Neuroplasticity Tim Klein, Beth Kendall, & Theresa Tougas St. Catherine University May 22, 2019 LIVED EXPERIENCE OF SELF-DIRECTED NEUROPLASTICITY ii Acknowledgements We would like to collectively express heartfelt gratitude to the many people who made this project possible. We begin with our teacher and advisor, Dr. Carol Geisler, associate professor at St. Catherine University. Thank you, Carol, for your profound passion and wisdom, sharp insights, and spirited sense of humor. Because of you, we continue to see possibilities.
    [Show full text]