Monday, April 16th, 2018

9:00 am9:45 am

Over the past twenty years, neuroscience research on reward-based learning has converged on a canonical model, under which the neurotransmitter dopamine 'stamps in' associations between situations, actions and rewards by modulating the strength of synaptic connections between neurons. However, a growing number of recent findings have placed this standard model under strain. I will present computational work in which we have drawn on recent advances in artificial intelligence to introduce a new theory of reward-based learning. Here, the dopamine system trains another part of the brain, the prefrontal cortex, to operate as its own free-standing learning system. This new perspective accommodates the findings that motivated the standard model, but also deals gracefully with a wider range of observations, providing a fresh foundation for future research.

Session Chair: Christos Papadimitriou

9:45 am10:30 am

When learning to make appropriate choices in different situations, humans can use multiple strategies in parallel, including working memory and reinforcement learning. Working memory allows very fast learning, but is cognitively effortful as well as limited in how much information can be retained, and for how long. Reinforcement learning has broader scope, but is slower and more incremental. Here, we investigate whether these two functions are independent in their computations and simply compete for choice, or if they interact at a deeper level. In multiple independent games, participants learned to select actions for varying numbers of new stimuli. When learning a low number of associations, performance was near optimal, indicating working memory use; however, with increasing number of items to learn, performance gradually decayed to a more incremental learning profile, as expected with from slower reinforcement learning mechanism. We will show evidence from fMRI, EEG and behavioral studies that the working memory process influences reinforcement learning computations, and specifically the update of estimated values with reward prediction errors. Indeed, this value update was surprisingly weakened in the easier conditions where performance was best. We will use computational modeling to show evidence that this is compatible with a competitive or cooperative interaction between working memory and reinforcement learning, but not with assuming that they are independent. We will then show preliminary evidence supporting the cooperative hypothesis, whereby working memory contributes expectations to the computation of the reward prediction error.

Session Chair: Christos Papadimitriou

11:00 am11:45 am

The ability to predict future states of the world is essential for planning behavior, and it is arguably a central pillar of intelligence.  In the field of sensory neuroscience, "predictive coding" -- the notion that circuits in cerebral actively predict their own activity -- has been an influential theoretical framework for understanding visual cortex.  In my talk, I will bring together the idea of predictive coding with modern tools of machine learning to build practical, working vision models that predict their inputs in both space and time. These networks learn to predict future frames in a video sequence, with each layer in the network making local predictions and only forwarding deviations from those predictions to subsequent network layers. We show that these networks are able to robustly learn to predict the movement of synthetic (rendered) objects, and that in doing so, the networks learn internal representations that are useful for decoding latent object parameters (e.g. pose) that support object recognition with fewer training views. We also show that these networks can scale to complex natural image streams (car-mounted camera videos), capturing key aspects of both egocentric movement and the movement of objects in the visual scene, and generalizing well across video datasets. These results suggest that prediction represents a powerful framework for unsupervised learning, allowing for implicit learning of object and scene structure.  At the same time, we find that models trained for prediction also recapitulate a wide variety of findings in neuroscience and psychology, providing a touch point between deep learning and empirical neuroscience data.

Session Chair: Christos Papadimitriou

TBA
11:45 am12:30 pm

No abstract available.

Session Chair: Christos Papadimitriou

2:30 pm3:15 pm
Speaker: Asja Fischer, University of Bonn

In recent years (deep) neural networks got the most prominent models for supervised machine learning tasks. They are usually trained based on stochastic gradient descent where backpropagation is used for the gradient calculation. While this leads to efficient training, it is not very plausible from a biological perspective.

We show that Langevin Markov chain Monte Carlo inference in an energy-based model with latent variables has the property that the early steps of inference, starting from a stationary point, correspond to propagating error gradients into internal layers, similar to backpropagation. Backpropagated error gradients correspond to temporal derivatives with respect to the activation of hidden units. These lead to a weight update proportional to the product of the presynaptic firing rate and the temporal rate of change of the postsynaptic firing rate. Simulations and a theoretical argument suggest that this rate-based update rule is consistent with those associated with spike-timing-dependent plasticity. These ideas could be an element of a theory for explaining how brains perform credit assignment in deep hierarchies as efficiently as backpropagation does, with neural computation corresponding to both approximate inference in continuous-valued latent variables and error backpropagation, at the same time.

Session Chair: Murray Sherman

3:15 pm4:00 pm
Speaker: Veronica Galvin, Yale University

The neural circuits underlying brain function across cortical regions have expanded and changed greatly throughout evolution.  The most striking of these changes is in the prefrontal cortex, where even across primate species there has been dramatic expansion of dendritic arbor, spine density, circuit connections, and transmitter and receptor actions in layer III pyramidal cells through evolution.  These changes in circuit architecture and function are critical to our understanding of brain function and cognition.  One critical process reliant on prefrontal cortex is working memory, or holding information “in mind” in the absence of sensory input.  This talk will discuss what we know of the anatomy and physiology of the circuitry underlying spatial working memory, which are known to function differently from other cortical regions such as primary visual cortex.  Computational modeling has enhanced our understanding and guided future studies of these circuits, specifically in their dependence on activation of postsynaptic NMDA receptors containing NR2B subunits.  Recent work has also shown a transition in these circuits away from a gating role of AMPA receptors, which instead rely on cholinergic actions on nicotinic alpha7 receptors.  Additional data support unique regulatory actions of cAMP, gating ion channels such as HCN and KCNQ, and this regulation shows opposite actions to primary sensory cortices. Understanding the complex and unique actions of various receptors and ion channels on these complex circuits is a key area of research to inform more accurate computational models of prefrontal cortex and improve our understanding of the brain.

Session Chair: Murray Sherman

Tuesday, April 17th, 2018

TBA
9:00 am9:45 am

No abstract available.

Session Chair: Veronica Galvin

9:45 am10:30 am
Speaker: Jeff Hawkins, Numenta

In this talk I will propose that the neocortex learns models of objects using the same methods that the entorhinal cortex uses to map environments. I will propose that each cortical column contains cells that are equivalent to grid cells. These cells represent the location of sensor patches relative to objects in the world. As we move our sensors, the location of the sensor is paired with sensory input to learn the structure of objects. I will explore the evidence for this hypothesis, propose specific cellular mechanisms that the hypothesis requires, and suggest how the hypothesis could be tested.

“A Theory of How Columns in the Neocortex Enable Learning the Structure of the World”; Hawkins, Ahmad, Cui; 2017
“Place Cells, Grid Cells, and the Brain’s Spatial Representation System”; Moser, Kropff, Moser; 2008
“Evidence for grid cells in a human memory network”; Doeller, Barry, Burgess; 2010

11:00 am11:45 am

Synaptic connections between neurons in the brain are dynamic because of continuously ongoing spine dynamics, axonal sprouting, and other processes. In fact, it was recently shown that the spontaneous synapse-autonomous component of spine dynamics is at least as large as the component that depends on the history of pre- and postsynaptic neural activity. These data are inconsistent with common models for network plasticity, and raise the questions how neural circuits can maintain a stable computational function in spite of these continuously ongoing processes, and what functional uses these ongoing processes might have. I will present a general theoretical framework that allows us to answer these questions. Our results indicate that spontaneous synapse-autonomous processes, in combination with reward signals such as dopamine, can explain the capability of networks of neurons in the brain to configure themselves for specific computational tasks, and to compensate automatically for later changes in the network or task.
On a more general level, the novel framework allows us to analyze synaptic plasticity and network rewiring from the viewpoint of stochastic optimization and sampling methods. As an example, I will show that the framework is also applicable to artificial neural networks, which provides as a side-effect new and effective brain-inspired methods for deep learning. 

References:

Kappel, D., Habenschuss, S., Legenstein, R., & Maass, W. (2015). Network plasticity as Bayesian inference. PLoS computational biology, 11(11), e1004485.

Kappel, D., Legenstein, R., Habenschuss, S., Hsieh, M., & Maass, W. (2018). A dynamic connectome supports the emergence of stable computational function of neural circuits through reward-based learning. arXiv preprint arXiv:1704.04238.


Bellec, G., Kappel, D., Maass, W., & Legenstein, R. (2017). Deep Rewiring: Training very sparse deep networks. arXiv preprint arXiv:1711.05136.

11:45 am12:30 pm

There has been rapid progress in the application of machine learning to  difficult problems such as: voice and image recognition, playing video games from raw-pixels, controlling high-dimensional motor systems, and winning at the games of Go, Chess and Poker. These recent advances have been made possible by employing the backpropagation-of-error algorithm. This algorithm enables the delivery of detailed error feedback to adjust synaptic weights, which means that even large networks can be effectively trained. Whether or not the brain employs similar deep learning algorithms remains contentious; how it might do so remains a mystery. I will begin by reviewing advances in deep reinforcement learning that highlight the importance of backprop for effectively learning complex behaviours. Then I will describe recent neuroscience evidence that suggests an increasingly complex picture of the neuron. This picture emphasizes the importance of electrotonically segregated compartments and the computational role of dendrites. Taken together, these findings suggest new ways that deep learning algorithms might be implemented in cortical networks in the brain.

2:30 pm3:15 pm

This talk will describe my recent work with Cameron Musco and Merav Parter, on studying neural networks from the perspective of the field of Distributed Algorithms.   In our project, we aim both to obtain interesting, elegant theoretical results, and also to draw relevant biological conclusions.

We base our work on simple Stochastic Spiking Neural Network (SSN) models, in which probabilistic neural components are organized into weighted directed graphs and execute in a synchronized fashion.  Our model captures the spiking behavior observed in real neural networks and reflects the widely accepted notion that spike responses, and neural computation in general, are inherently stochastic.  In most of our work so far, we have considered static networks, but the model would allow us to also consider learning by means of weight adjustments.

Specifically, we consider the implementation of various algorithmic primitives using stochastic SNNs.  We first consider a basic symmetry-breaking task that has been well studied in the computational neuroscience community:  the Winner-Take-All  (WTA)  problem.  WTA is believed to serve as a basic building block for many other tasks, such as learning, pattern recognition, and clustering.  In a simple version of the problem, we are given neurons with identical firing rates, and want to select a distinguished one.  Our main contribution is the explicit construction of a simple and efficient WTA network containing only two inhibitory neurons; our construction uses the stochastic behavior of SNNs in an essential way.  We give a complete proof of correctness and analysis of convergence time, using distributed algorithms proof methods.  In related results, we give an optimization of the simple two-inhibitor network that achieves better convergence time at the cost of more inhibitory neurons.  We also give lower bound results that show inherent limitations on the convergence time achievable with small numbers of inhibitory neurons.

We also consider the use of stochastic behavior in neural algorithms for Similarity Testing (ST).   In this problem, the network is supposed to distinguish, with high reliability, between input vectors that are identical and input vectors that are significantly different.  We construct a compact stochastic network that solves the ST problem, based on randomly sampling positions in the vectors.  At the heart of our solution is the design of a compact and fast-converging neural Random Access Memory (neuro-RAM)  mechanism.

In this talk, I will describe our SNN model and our work on Winner-Take-All, in some detail.  I will also summarize our work on Similarity Testing, discuss some important general issues such as compositionality, and suggest directions for future work.

3:15 pm4:00 pm

I will sketch a new approach for understanding computation and learning in recurrent networks of spiking neurons. This approach takes into account that networks of neurons in the brain have undergone long evolutionary processes and prior learning experiences before they encounter a new learning task. Hence they work in a quite different setting than most of our models, which are expected to learn from scratch.

Learning-to-learn (L2L) methods in machine learning, such as (Hochreiter et al., 2001) and (Wang et al., 2016), have started to address this inadequacy for artificial neural networks. I will show that some of these methods can be ported to biologically quite realistic networks of spiking neurons, provided one has spike-based modules for working memory that can emulate the function of Long Short-term memory (LSTM) modules from artificial neural networks. The inclusion of biologically more realistic neuron models, in combination with a suitable network architecture, turns out to provide such spike-based LSTM modules. The resulting application of L2L methods to networks of spiking neurons shows that they are able to learn in previously not expected ways. For example, they can learn complex input/output behaviors, for which one would expect that nonlocal learning algorithms such as backprop is required, with simple local learning rules.

The new spike-based LSTM modules are also of interest from two other perspectives. First, they  provide new ways of modeling computational operations of brains which require a working memory. Secondly, they enable networks of spiking neurons to attain some of the astounding computational capabilities that have been demonstrated during the last few years for recurrent networks of LSTM-modules.

Background information can be found in:
---Hochreiter, S., Younger, A. S., & Conwell, P. R. (2001, August). Learning to learn using gradient descent. In International Conference on Artificial Neural Networks (pp. 87-94). Springer, Berlin, Heidelberg.
---Wang, J. X., Kurth-Nelson, Z., Tirumala, D., Soyer, H., Leibo, J. Z., Munos, R., ... & Botvinick, M. (2016). Learning to reinforcement learn. arXiv preprint arXiv:1611.05763.
---Bellec, G., Salaj, D., Subramoney, A., Legenstein, R., Maass, W. (2018) Long short-term memory in networks of spiking neurons. Arxiv 2018

Wednesday, April 18th, 2018

9:00 am9:45 am

Visual stimuli evoke activity in visual cortical neuronal populations. Neuronal activity can be selectively modulated by particular visual stimulus parameters, such as the direction of a moving bar of light, resulting in well-defined trial averaged tuning properties. However, a large number of neurons in visual cortex remain unmodulated by any given stimulus parameter, and the role of this untuned population is not well understood. Here, we use two-photon calcium imaging to record, in an unbiased manner, from large populations of layer 2/3 excitatory neurons in mouse primary visual cortex to describe co-varying activity on single trials in neuronal populations consisting of both tuned and untuned neurons. Specifically, we summarize pairwise covariability with an asymmetric partial correlation coefficient, allowing us to analyze the resultant population correlation structure, or functional network, with graph theory. Using the graph neighbors of a neuron, we find that the local population, including both tuned and untuned neurons, are able to predict individual neuron activity on a moment to moment basis and while also recapitulating tuning properties of tuned neurons. Variance explained in total population activity scales with the number of neurons imaged, suggesting larger sample sizes are required to fully capture local network interactions. We also find that a specific functional triplet motif in the graph results in the best predictions, suggesting a signature of informative correlations in these populations. In summary, we show that unbiased sampling of the local population can explain single trial response variability as well as trial-averaged tuning properties in V1, and the ability to predict responses is tied to the occurrence of a functional triplet motif.

9:45 am10:30 am

The sparse coding model has been shown to provide a good account of neural response properties at early stages of sensory processing.  However, despite several promising efforts [1,2,3] it is still unclear how to exploit the structure in a sparse code for learning higher-order structure at later stages of processing.  Here I shall argue that the key lies in understanding how continuous transformations in the signal space are expressed in the elements of a sparse code, and in deriving the proper computations that disentangle these transformations from the underlying invariances.  I shall present a new signal representation framework, called the sparse manifold transform, that exploits temporally-persistent structure in the input (similar to slow feature analysis) in order to turn non-linear transformations in the signal space into linear interpolations in a representational embedding space.  The SMT thus provides a way to progressively flatten manifolds [4], allowing higher forms of structure to be learned at each higher stage of processing.  The SMT also provides a principled way to derive the pooling layers commonly used in deep networks, and since the transform is approximately invertible, dictionary elements learned at any level in the hierarchy may be directly visualized.  Possible neural substrates and mechanisms of SMT shall be discussed.

With Yubei Chen.

[1] Karklin, Y., & Lewicki, M. S. (2003). Learning higher-order structures in natural images. Network: Computation in Neural Systems, 14(3), 483-499.

[2] Hosoya, H., & Hyvärinen, A. (2015). A hierarchical statistical model of natural images explains tuning properties in V2. Journal of Neuroscience, 35(29), 10412-10428.

[3] Le, Q. V. (2012). Building high-level features using large scale unsupervised learning. In Proceedings of the 29th International Conference on Machine Learning.
https://arxiv.org/abs/1112.6209

[4] DiCarlo, J.J., & Cox, D. D. (2007)  Untangling invariant object recognition.  Trends in Cognitive Sciences, 11: 333-341.

11:00 am11:45 am

Building a microscopically accurate model of a system as complex as a brain or even a single cell is hardly possible, and arguably not very useful. Can we use modern machine learning approaches to automate finding predictive phenomenological dynamical models of time series measured in experiments? I will discuss our approach to the problem, implemented as a software Package SirIsaac. I will show how the method performs in various synthetic test cases, as well as on building (and interpreting) a dynamical model of a reflexive escape from a painful stimulus by C. elegans.

11:45 am12:30 pm

In attempts to understand the functional circuits in cortex and thalamus, it is important to recognize that components of the predominant excitatory circuits, which use glutamate as a neurotransmitter, are not functionally homogeneous, as they are often mistakenly represented. Instead, these glutamatergic pathways can be divided into two distinct classes: Driver, which carries the main information between cells, and Modulator, which modifies how driver inputs function. Identifying which glutamatergic pathways are which is an important prerequisite to further understanding of circuit function. Examples of how this can be applied to thalamic and cortical circuitry will be discussed.

2:30 pm3:15 pm

Recent understanding of our nervous system reveals that sensory organs are optimized for efficient uptake of information from the environment to the brain. While bulk of the energy consumed in this process is to generate asynchronous action potentials for signal communication, evolutionary pressure has further optimized on this mode of communication within the brain for efficient representation and processing of information. We will explore this link between energy and information to reveal some of the mechanisms, circuits and energy efficient codes in the brain. The implications of this link will be important to consider for the design of neuromorphic systems of the future.

3:15 pm4:00 pm

A neural system can flexibly perform many tasks, the underlying mechanism is unknown. Here, we trained a single recurrent network model, using a machine learning method, to perform 20 cognitive tasks that may involve working memory, decision-making, categorization and inhibitory control. We found that after training the emerging task representations are organized in the form of clustering of recurrent units. Moreover, we introduce a measure to quantify single-unit neural relationships between 190 pairs of tasks, and report five distinct types of such relationship that can be tested experimentally. Surprisingly, our network developed compositionality of task representations, a critical feature for cognitive flexibility, thereby one task can be instructed by combining representations for other tasks. Finally, we demonstrate how the network could learn multiple tasks sequentially. This work provides a computational platform to investigate neural representations of many cognitive tasks, and suggests new research directions at the interface between neuroscience and AI.

Thursday, April 19th, 2018

9:00 am9:45 pm

Many forms of learning are guided by rewards and punishments. Humans and animals can be trained in complex tasks that consist of multiple epochs by the appropriate choice of reward contingencies. It is not well understood how association cortices learn to link sensory stimuli and memory representations to motor programs during reinforcement learning. Furthermore, there are many processing steps between the sensory cortices and the motor cortex. The learning rules for training such deep biological networks are only partially understood. I will outline a theory that explains how deep brain networks can learn a large variety of tasks when only stimuli and reward contingencies are varied, and present neurophysiological evidence in support of this theory.

The aim of the proposed architecture is to learn action values, i.e. the value of an action in a specific situation. I will demonstrate how networks can learn action values when they utilizes an ‘attentional’ feedback signal from motor cortex back to association cortex that “tags” synapses that should change to improve behavior. The resulting learning rule can train a simple neural network in many tasks that are in use in neurophysiology, including (1) delayed saccade tasks; (2) memory saccade tasks; (3) saccade-antisaccade tasks; (4) decision making tasks; and (5) classification tasks.

The proposed theory predicts that neurons at intermediate levels acquire visual responses and memory responses during training that resemble the tuning of neurons in association areas of the cerebral cortex of monkeys. The learning rule predicts that action values influence neuronal activity in sensory areas, as an attentional feedback effect. It is encouraging that insights from molecular, cellular and systems neuroscience can now be combined with insights from theories of reinforcement learning and deep artificial networks to develop a unified framework for learning in the brain. 

11:00 am11:45 am

We investigate neural circuits in the exacting setting that (i) the acquisition of a piece of knowledge can occur from a single interaction, (ii) the result of each such interaction is a rapidly evaluatable subcircuit, (iii) hundreds of thousands of such subcircuits can be acquired in sequence without substantially degrading the earlier ones, and (iv) recall can be in the form of a rapid evaluation of a composition of subcircuits that have been so acquired at arbitrary different earlier times.

We develop a complexity theory, in terms of asymptotically matching upper and lower bounds, on the capacity of a neural network for executing, in this setting, the following action, which we call association: Each action sets up a subcircuit so that the excitation of a chosen set of neurons A will in future cause the excitation of another chosen set B. A succession of experiences, possibly over a lifetime, results in the realization of a complex set of subcircuits. The composability requirement constrains the model to ensure that, for each association as realized by a subcircuit, the excitation in the triggering set of neurons A is quantitatively similar to that in the triggered set B, and also that the unintended excitation in the rest of the system is negligible. These requirements ensure that chains of associations can be triggered.

We first analyze what we call the Basic Mechanism, which uses only direct connections between neurons in the triggering set A and the target set B. We consider random networks of n neurons with expected number d of connections to and from each. We show that in the composable context capacity growth is limited by d^2, a severe limitation if the network is sparse, as it is in cortex. We go on to study the Expansive Mechanism, that additionally uses intermediate relay neurons which have high synaptic weights. For this mechanism we show that the capacity can grow as dn, to within logarithmic factors. From these two results it follows that in the composable regime, for the realistic estimate of d being the square root of n, optimal superlinear capacity in terms of the neuron numbers can be realized by the Expansive Mechanism, instead of the linear order n to which the Basic Mechanism is limited.

Reference: L.G Valiant, Capacity of Neural Networks for Lifelong Learning of Composable Tasks,  Proc. 58th Annual IEEE Symposium on Foundations of Computer Science, October 15 - 17, 2017, Berkeley, California , 367-378 (2017).

11:45 am12:30 pm

Animals execute goal-directed behaviors despite the limited range and precision of their sensors. To cope, they explore environments and store memories maintaining estimates of important information that is not presently available. Recently, breathtaking progress has been made with artificial intelligence (AI) agents that learn to perform tasks from sensory input, even at a human level, by merging reinforcement learning (RL) algorithms with deep neural networks, and the excitement surrounding these results has led to the pursuit of related ideas as explanations of non-human animal learning. However, we demonstrate that contemporary RL algorithms are unable to solve simple tasks when enough information is concealed from the sensors of the agent, a property called ``partial observability''. An obvious requirement for handling partially observed tasks is access to extensive memory, but we show memory is not enough; it is critical that the right information be stored in the right format. We develop a model, the Memory, RL, and Inference Network (MERLIN), in which memory formation is guided by a process of predictive modeling. MERLIN breaks ground to solve new classes of tasks in 3D virtual reality environments for which partial observability is severe and memories must be maintained over long durations. Our model represents an advance in AI and, we hope, a useful conceptual framework for neuroscience, demonstrating a cognitive architecture that can solve canonical behavioral tasks in psychology and neurobiology without strong simplifying assumptions about the dimensionality of sensory input or the duration of experiences.

Work done in conjunction with Chia-Chun Hung, David Amos, Mehdi Mirza, Jack Rae, Arun Ahuja, Agnieszka Grabska-Barwinska, Piotr Mirowski, Timothy Lillicrap, and others at DeepMind.