Bulletin of the American Physical Society
APS March Meeting 2022
Volume 67, Number 3
Monday–Friday, March 14–18, 2022; Chicago
Session A03: Physics of Learning I: Natural systemsFocus Recordings Available
|
Hide Abstracts |
Sponsoring Units: DBIO GSNP Chair: Ilya Nemenman, Emory Room: McCormick Place W-176A |
Monday, March 14, 2022 8:00AM - 8:36AM |
A03.00001: A mathematical analysis of learning in both biological and artificial neural networks Invited Speaker: Surya Ganguli An exciting area of intellectual activity in this century may well revolve around a synthesis of machine learning, theoretical physics, and neuroscience. The unification of these fields will likely enable us to exploit the power of complex systems analysis, developed in theoretical physics and applied mathematics, to elucidate the design principles governing neural systems, both biological and artificial, and deploy these principles to develop better algorithms in machine learning. We will give several vignettes in this direction, including: (1) determining the best optimization problem to solve in order to learn regression in high dimensions; (2) finding exact solutions to the dynamics of generalization error in deep linear networks; (3) analyzing how neural networks can learn semantic concepts like infants. |
Monday, March 14, 2022 8:36AM - 8:48AM |
A03.00002: Intrinsic and parameter-less gain control in rate coding by spiking neurons Nirag Kadakia, Will Rosenbluth, Thierry Emonet Adaptation is a critical feature of neural systems, including individual neurons. In single neurons, contrast adaptation be mediated by slow processes such as a calcium influx. Here, we propose a biophysical mechanism for contrast adaptation that does not rely on changes in internal state. Our framework is motivated by observations of Drosophila olfactory receptor neurons to fluctuating stimuli: firing rates do not modulate smoothly, rather they switch more discontinuously between low and high ~40 Hz. In the language of dynamical systems, the system persistently crosses a bifurcation between spiking and quiescence. Typically, this system could only encode 1 bit of information, but we show that the conversion from spike events to a rate code encodes more than 1 bit. In addition, responses are contrast invariant: thus, bifurcation crossing amplifies small fluctuations, permitting rate codes that would otherwise be imperceptible. Such bifurcation-induced gain control is a general property of neurons with various bifurcation topologies. Our results suggest that the machinery of neuron spiking permits robust adaptation with high coding efficacy. |
Monday, March 14, 2022 8:48AM - 9:00AM |
A03.00003: Neural network model of nonlinear Bayesian learning Camilla Li, Ilya M Nemenman Animals learn sensorimotor tasks by performing actions, getting sensory feedback, and incorporating the feedback into their knowledge of the world to guide future actions. Previous work from our group revealed that this process, in the context of pitch compensation in songbirds learning to sing, can be modeled by a Bayesian filter with non-Gaussian probability distributions. How such computations can be implemented using neural networks remains unknown. Inspired by line attractor neural network models, we develop a neural network with stochastic dynamics, which stores information about optimal actions in the positions of activity bumps, and represents the growing uncertainty about the actions in the diffusion of these bumps with time. The model explains salient features of experiments, such as a dynamically changing variability of actions, a bimodal distribution of actions, jumps and continuous responses to external perturbation, and a common disregard for sensory feedback that contradicts prior expectations. |
Monday, March 14, 2022 9:00AM - 9:12AM |
A03.00004: Desynchronous and decentralized learning In a physical network Jacob F Wycoff, Sam J Dillavou, Menachem Stern, Andrea J Liu, Douglas J Durian Previous work has realized what we call a physics-driven learning network: an electrical network of variable resistors capable of learning a variety of tasks using simple local rules and natural power minimization [1]. Like biological neuron networks, this network performs distributed computation using exclusively local information, and thus is scalable and robust to damage. However, it uses a global clock to update each edge simultaneously. Here, we take the system one step closer to its biological inspiration by relaxing the global clock requirement, allowing individual edges of the network to update randomly and independently. We find that this stochasticity does not impair performance, and can even improve the learning error for some tasks. This effect can be understood by analogy to Stochastic Gradient Descent. Our results suggest that our approach can not only be implemented in metamaterials or other sensors where a central clock is untenable, but can also improve learning. |
Monday, March 14, 2022 9:12AM - 9:24AM |
A03.00005: Merging Kalman Filtering with the information bottleneck technique for optimal state inference Vedant Sachdeva, Thierry Mora, Arvind Murugan, Stephanie E Palmer, Aleksandra M Walczak Living systems must make estimates of a state of a stimulus in order to make decisions relevant for survival. One approach living systems can take to constructing these estimates involves leveraging past state estimates and present sensory responses in a Kalman filter. However, unlike in engineering, living systems are capable of changing their sensory response over evolutionary timescales, subject to constraints on metabolism, computing power, and other biologically relevant limitations. The optimal choice for a sensory model navigates the resource constraints, providing information about the current state, and providing information about future states for later inference. To determine how this sensory model should be constructed, we use the Information Bottleneck method. By connecting Kalman filter state inference with Information Bottleneck-based sensory models, we can demonstrate how biological systems can optimally take advantage of both memory and sensory information simultaneously. We show this in both one- and two-dimensional Gaussian stimuli. We present the analytic solution to one-dimensional stimuli, and observe several regimes for the sensory model as a function of correlation time and resource constraints. |
Monday, March 14, 2022 9:24AM - 9:36AM |
A03.00006: A theory of weight distribution-constrained learning Weishun Zhong, Ben Sorscher, Daniel D Lee, Haim I Sompolinsky Recent large-scale connectomics studies have provided precise insights into the excitatory/inhibitory identities of individual synapses, as well as the distribution of synaptic weights in the brain. Motivated by this, we developed a theory of learning in neural networks that incorporates both sign and distribution constraints. We found analytical solutions for both the capacity and generalization performance in perceptron, a basic feedforward model, and developed an SGD-based algorithm to find weights that satisfies these constraints. We further applied our results to the Hopfield network, a recurrent model, and demonstrated that heterogeneity in neural populations emerges from a global distribution constraint. |
Monday, March 14, 2022 9:36AM - 9:48AM |
A03.00007: Biophysical learning : Training flow networks via chemical signaling Vidyesh Rao Anisetti, Benjamin Scellier, Siddhartha Mishra, Jennifer M Schwarz Biological systems have the capability to modify their individual components to achieve task-specific global functionality. Consider, for example, slime mold modifying its structure to optimize for transport between food sources, brain modifying its synapses to improve behaviour or to compensate for loss of functionality in a damaged region. Our work explores how modifications in a physical system using locally available information can give rise to emergent global functionality. Inspired by how slime mold modifies its structure using chemical signals, we propose a mechanism to adjust the resistances of a flow network using chemical signaling. We demonstrate that such a mechanism can train a simple flow network to classify three species of iris flowers with significant accuracy by using the information of the sizes of the flower components as input, thus exhibiting 'machine learning-like' behavior. We also show that this mechanism optimizes a loss function by gradient descent. These simple biophysical learning systems may ultimately help us understand how more complex biological networks, such as the brain, exhibit emergent functionality. |
Monday, March 14, 2022 9:48AM - 10:00AM |
A03.00008: How physics-informed neural networks improve image segmentation performance Daniela Koch Many quantitative cell biology experiments depend on rapid and faithful segmentation (the computational identification) of cells in microscopy images. Recent developments in deep learning have greatly improved these algorithms; however, significant challenges remain, especially in the context of cells with unusual morphologies. A critical limitation of deep learning is the size and diversity of the training set, which often must be hand segmented by an expert. An alternative approach is the use of data augmentation, using a physics-informed neural network, to supplement training. We demonstrate that this approach greatly increases the performance of the segmentation algorithm. This is a promising example of a potentially widely applicable approach: the use of physics to constrain a deep learning model to increase performance at fixed training set size. |
Monday, March 14, 2022 10:00AM - 10:12AM |
A03.00009: Exploring the Landscape of Collective Modes Using IB-RG Adam G Kline The Renormalization Group (RG) is a set of theoretical techniques which interpolate between microscopic models and their macroscopic properties. However, RG typically relies on a definition of the collective modes of a system, organized from least to most relevant. In biological systems, this poses a challenge, since context and function shape collective behaviors in a way that is not known at the outset. In another domain, the information bottleneck approach (IB) describes the best probabilistic compression of an input variable X under a pre-defined “relevance” variable Y. Recent theoretical work has shown that IB can be imposed on RG, allowing one to define the notion of large-scale structure probed by the RG flow [1]. To understand what it means to choose a notion of scale, we explore the behavior of IB-RG flows as the relevance variable statistics are changed. We uncover nontrivial physics in terms of order parameters that vary with with our definition of Y. Our analysis may serve as a model for RG applied to biological systems, in particular neural networks, whose goals are multifaceted and whose collective behavior requires sensitivity to many features in the environment. |
Monday, March 14, 2022 10:12AM - 10:24AM |
A03.00010: Energy-Based Models Capture Pairwise and Higher-Order Interactions in Protein Sequence Data Peter Fields, Vudtiwat Ngampruetikorn, Stephanie E Palmer, David J Schwab Understanding protein structure, evolution and function requires reliable inference of interacting units in folded proteins. Here we present a unifying approach for inferring two of the most important structural units of proteins: pairwise contacts, and higher-order strongly correlated units, known as sectors. Our method is a hybrid energy-based model, combining a pairwise-energy term, as used in state-of-the-art Direct Coupling Analysis, and a Restricted Boltzmann Machine (RBM) term, meant to capture higher order interactions. We show that, when trained on data from a biologically-informed ground truth model, our algorithms can learn both the pairwise and higher-order structure and are robust to varying levels of undersampling and strength of interactions in the ground truth distribution. We carry out the analysis for 2-spin and 10-spin systems with Minimum Probability Flow and Ratio Matching algorithms, respectively. We comment on why the RBM is successful at modeling the higher-order interactions and why certain choices for hyperparameters (number of hidden units in the RBM, regularization strength) lend themselves to the model's feature detection capabilities. |
Monday, March 14, 2022 10:24AM - 10:36AM |
A03.00011: Learning in gene regulatory networks: dimensionality reduction by master regulators Naama Brenner Under stressful and unforeseen challenge, cells can harness their internal complexity and plasticity to acquire novel adaptive phenotypes. Such exploratory adaptation is a primitive analog of learning; its dynamics and properties are not well understood. I will present a computational framework to describe such adaptation inspired by learning models of the brain. A random network model of gene regulation shows the feasibility of exploratory adaptation based on purely stochastic dynamics and stress sensing. Convergence in high-dimensional gene expression space is non-universal and depends on network properties: Specifically, convergence is promoted by heterogeneous connections with outgoing hubs – “master regulators”, a known feature of gene networks. We construct a coarse-grained model for understanding the role of hubs in the search process, that maps the problem onto suppression of chaotic network activity by an external drive. The phase transition in that problem sheds light on the role of master regulators in reducing search dimensionality. |
Monday, March 14, 2022 10:36AM - 10:48AM |
A03.00012: Memory formation in adaptive networks Komal Bhattacharyya, David Zwicker, Karen Alim Continuous adaptation of networks like our vasculature ensures optimal network performance when challenged with changing loads. Here, we show that adaptation dynamics allow a network to memorize the position of an applied load within its network morphology. We identify that the irreversible dynamics of vanishing network links encode memory. Our analytical theory successfully predicts the role of all system parameters during memory formation, including parameter values which prevent memory formation. We thus provide an analytically tractable theory of memory formation in disordered systems. |
Monday, March 14, 2022 10:48AM - 11:00AM Withdrawn |
A03.00013: The route to chaos of reinforcement learning in routing networks Jakub Bielawski, Thiparat Chotibut, Fryderyk Falniowski, Michal Misiurewicz, Georgios Piliouras Without a central authority to manage traffic in a routing network, can a society consisting of uncoordinated, self-interested agents automatically reach the optimal traffic flow that minimizes the travel time of the society? For decades, game theory framework has been applied to study a set of stable strategies (Nash equilibria) that self-interested agents might adopt in these routing games. However, such framework focuses on the identification of equilibria, rather than on whether such equilibria are attainable if the agents individually and strategically learn from the history. In this work, we show that collective behaviors of reinforcement learning agents can be driven away from the Nash equilibrium. In fact, a period-doubling bifurcation route to chaos naturally emerges as the traffic load increases. Interestingly, even when the collective behaviors are chaotic (in the Li-Yorke sense), their ergodic average still coincides exactly with the Nash equilibrium. Lastly, we report the numerical evidence of the Feigenbaum's universality class in our non-unimodal map. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700