Abstracts

Abstracts for program presentations

Detailed workshop information is shared to registered participants.

Tuesday Aug 4: Learning to Control

Rodolphe Sepulchre (Cambridge, UK): Learning to control artificial neurons with neuromodulation.

TBA

Yorie Nakahira (Carnegie Mellon Univ, Pittsburgh): Diversity-enabled sweet spots in layered architectures and speed-accuracy trade-offs in sensorimotor control.

Joint work with Quanying Liu, Terry Sejnowski, and John Doyle

Abstract: Nervous systems sense, communicate, compute, and actuate movement using distributed components with trade-offs in speed, accuracy, sparsity, noise, and saturation. The resulting control can achieve fast, accurate, and robust performance due to a highly effective layered control architecture. But there is no theory explaining the effectiveness of layered control architectures that connects speed-accuracy trade-offs (SATs) in neurophysiology to the resulting SATs in sensorimotor control. We developed a Telluride-inspired mountain biking driving task and manipulate the time delays and accuracy of the control input from the wheel as a surrogate for manipulating the characteristics of neurons in the control loop. The speed-accuracy trade-offs observed in drivers' behavior motivated a theoretical framework consisting of layers of processing in the control loop with different time scales and several levels of hardware organization in the brain with diverse speeds and accuracy. This framework characterizes how the sensorimotor control SATs are constrained by the hardware SATs of the sensory and motor neurons. These results demonstrate that an appropriate diversity in the properties of neurons and muscles across layers and within levels helps to create systems that are both fast and accurate despite being built from components that are individually slow or inaccurate. These ``diversity-enabled sweet spots'' explain the ubiquity of heterogeneity in the sizes of axons within a nerve and the resulting superior performance of sensorimotor control.

Martina Poletti (U Rochester): The synergy between oculomotor behavior and visual perception​.

Abstract: Outside the foveola, visual acuity and other visual functions gradually deteriorate with increasing eccentricity. Humans compensate for these limitations by relying on a tight link between perception and action; rapid gaze shifts (saccades) occur 2-3 times every second, separating brief “fixation” intervals in which visual information is acquired and processed. During fixation, however, the eye is not immobile. Small eye movements incessantly shift the image on the retina even when the attended stimulus is already foveated, suggesting a much deeper coupling between visual functions and oculomotor activity. By many considered a nuisance, fixational eye movements have long attracted the attention of researchers in visual perception and motor control. Recent technical advances have greatly improved investigation of the sensory-motor interaction occurring at fixation. Here, I will review evidence showing that fixational eye movements are not simply the outcome of limits in oculomotor control, as once assumed, but finely controlled behaviors that serve important visual functions. This body of work indicates that fine spatial vision critically relies on precise fixational eye movements.

Wednesday Aug 5: Cognition and Reinforcement Learning

Roy Fox (UC Irvine): Structured control as inference

The duality between optimal control and Bayesian inference has been influential in designing feedback control systems in which perception and action are separable. In recent years, following progress in approximate inference, it was realized that this duality also holds in approximation and is applicable to the complete perception–action cycle, lending solid probabilistic interpretation to several successful algorithms for planning and learning in dynamical systems. We extend this framework by noting two useful capabilities of variational inference methods. First, the ability to infer latent variables can go beyond the interaction variables, namely the observations and actions, and can discover useful structure of the agent's internal memory process. Second, variational inference allows the proposal model to be different from the generative model, which enables the extraction of an acausal learning signal that is unused in previous work. We demonstrate these benefits in hierarchical imitation learning of control programs.

Andrea Stocco (UW): Higher-level cognitive functions and the reptilian brain: How RL computations have pervasive signature effects in language, attention, control, and fluid intelligence.

The basal ganglia (BG) are an evolutionarily old part of the brain that is generally understood to perform reinforcement learning-type computations. While their role of the BG in learning and decision making is well-documented expected, in this talk I will show that basal ganglia computations are pervasive across cognitive domains and signatures of their activity can be traced in higher-level cognitive functions such as attention, executive functions, language production and comprehension, and even analogical reasoning and fluid intelligence. Taken together, these results paint a new picture of how the basal ganglia fit within the brain's architecture: in this view, rather than being a specific functional component, the basal ganglia are the foundation upon which more complex cognitive functions are layered and built upon.

Terry Stewart (NRC/UWaterloo): Representing Space and Time in Neurons

This talk looks at a variety of new results from my lab using neurons to represent continuous space and continuous time. In particular, it turns out that it is possible to represent locations of objects in unbounded spaces, and this approach works to represent spatial areas and maps well onto grid cells in the brain. For representing time, we can also derive an optimized recurrent network (similar to a reservoir) that efficiently stores the past information about the input. This not only corresponds well with time cells in the brain, but also turns out to be a state-of-the-art solution to machine learning problems, out-performing systems such as LSTMs on some problems.

Benjamin Scellier (MILA, Montréal): Training nonlinear resistive networks with equilibrium propagation

Abstract: Equilibrium propagation (EqProp) is an algorithm to train neural networks by gradient descent, using a single computational circuit (both for inference and training) and a local learning rule. Specifically, EqProp applies to energy-based neural networks, a class of models relying on equilibrium states. Recent work has shown that nonlinear resistive networks satisfy a variational principle and thus can implement energy-based neural networks efficiently. As a result, EqProp provides a method to train end-to-end analog networks by using Kirchhoff's laws as first principles.


Thursday Aug 6: Neuroscience and Machine Learning

Fritz Sommer (U. Berkeley): Computing with Rhythms and Spikes

Abstract: Information coding by precise timing of spikes can be faster and more energy efficient than traditional rate coding. However, spike-timing codes are often brittle, which so far has limited their use in neuromorphic computing. Experimental neuroscience suggests that coherent rhythms in neuronal membrane currents, manifested in temporal structure of local field potentials and electroencephalograms, carry detailed behavioral information and thus could be a critical component of reliable spiking computations. I will describe recent efforts to build a theory of how to compute with rhythms and spikes.

I will describe work (Frady & Sommer, PNAS, 2019), proposing a type of attractor neural network in complex state space, threshold phasor associative memory (TPAM), that can be leveraged to construct spiking neural networks with robust computational properties. Further, I will outline recent unpublished work on phasor theory that can describe interesting types of computation and communication in spiking neural networks, in particular, vector-symbolic cognitive reasoning, and flexible information routing based on the communication through coherence postulate. Finally, I will summarize the implications of the described theory on understanding phenomena in neuroscience, and as a framework for designing algorithms for emerging neuromorphic devices.


Wolfgang Maass and Anand Subramoney (TUG): New learning methods for recurrent networks of spiking neurons

Abstract: On-chip learning capability is obviously an essential need for neuromorphic hardware. Since most of this hardware is spike-based, and less suitable for efficient implementations of CNNs or other feedforward networks, chip-friendly online learning methods for recurrent networks of spiking neurons (RSNNs) are most urgently needed. There exist already many methods for that, but we are lacking methods that enable us to approximate gradient descent of rather arbitrary loss functions, which we need to port key achievements of modern Machine Learning and AI into NMH. We present e-prop as one possible method for that, and show in particular that how Deep RL can be ported in this way to NMH. We also present a recent variant of that, natural e-prop, that enables one-shot learning for RSNNs.

Details can be found in

----G. Bellec, F. Scherr, A. Subramoney, E. Hajek, D. Salaj, R. Legenstein, and W. Maass. A solution to the learning dilemma for recurrent networks of spiking neurons. Nature Communications (July 2020).

----F. Scherr, C. Stoeckl, and W. Maass. One-shot learning with spiking neural networks. bioRxiv, 2020

Claire Pelofi (NYU): Electrophysiological markers of timely cognitive processes: when to use EEG signals

Abstract: In 2024, Electroencephalography (EEG) will be celebrating one century of good offices to the field of human cognition and neuroscience. By providing a temporally precise (10 ms) signal reflecting electrical activity emitted by thousands of millions of neurons captured by 32-128 electrodes, EEG has undoubtedly contributed to the understanding of the human mind like no other neuroimaging technique. From oscillatory activity findings, to ERP studies and finishing with the most cutting-edge decoding approach, we will here span the many assets EEG techniques can offer for the study of human cognition. Focusing mostly on auditory perception at large, we will examine how those different EEG paradigms can enlighten our understanding of questions such as auditory object formation, syntax and regularities encoding, or attention tracking in eco-valid experimental set-ups.

Friday Aug 7: Technology and Challenges

Xin Wang (Cerebras): Wafer-scale neural computation

We introduce the Cerebras CS-1, a powerful and efficient computer system built to accelerate deep learning compute at scale in the datacenter. At its heart, CS-1 system is powered by Cerebras' Wafer-Scale Engine (WSE), the world's largest chip and first trillion-transistor processor. The WSE combines massive computing resources with high speed memory and communication to provide acceleration orders of magnitude beyond traditional, general purpose processors. The WSE has 400,000 programmable cores connected by a high bandwidth, low latency, configurable on-chip interconnect and fast, local memory--a non-Von Neumann architecture offering unprecedented opportunities for biologically inspired algorithm research. Recent progress in deep learning and its application to neuromorphic engineering suggests that the CS-1 could be a valuable tool for DNN research and training for models that could be deployed to run on emergent neuromorphic devices.

Mike Davies (Intel): Advancing neuromorphic computing from lab to mainstream applications

Abstract: Most if not all Telluride attendees are already familiar with Intel's Loihi research chip and neuromorphic computing program. This update describes what's new in the past year, namely a growing list of systems, software, algorithmic capabilities, and most importantly quantitative results showing compelling gains compared to conventional architectures. These results point to a roadmap of disruption spanning edge to datacenter computing applications that require low latency, low power, and adaptive processing.