Neural Theory
While neuroscientists have increasingly powerful deep learning models that predict neural responses, it is not clear that these models are correspondingly increasing our understanding of what neurons are actually doing. In this session, we will take a more mechanistic approach to understanding how networks of neurons afford complex computations, both by considering mechanistic neural models along with mathematical theories that say how neurons should behave and crucially why they behave that way.
Session Chairs
Dr James Whittington (University of Oxford; Stanford University)
Yasmine Ayman (Harvard)
Keynote Talks
Professor Christine Constantinople (New York University)
Invited Talks
Dr Juan Gallego (Imperial)
Professor Vijay Namdoodiri (Weill Institute for Neurosciences, UCSF)
Professor Annegret Falkner (Princeton Neuroscience Institute)
Dr Francesca Migancco (Princeton University)
Dr Valeria Fascianelli (Columbia)
Dr Chris Hillar (New Theory AI)
Spotlight Talks
Jesseba Fernando (Northeastern University)
Albert Albesa Gonzalez (Imperial College London)
Dr Alexander Rivkind (Cambridge)
Joseph Warren (Sainsbury Wellcome Centre - UCL)
Samuel Liebana (UCL)
Keynotes
NYU
Neural Circuit Mechanisms of Value-Based Decision-Making
The value of the environment determines animals’ motivational states and sets expectations for error-based learning. But how are values computed? We developed a novel temporal wagering task with latent structure, and used high-throughput behavioral training to obtain well-powered behavioral datasets from hundreds of rats that learned the structure of the task. We found that rats use distinct value computations for sequential decisions within single trials. Moreover, these sequential decisions are supported by different brain regions, suggesting that distinct neural circuits support specific types of value computations. I will discuss our ongoing efforts to delineate how distributed circuits in the orbitofrontal cortex and striatum coordinate complex value-based decisions.
Invited Speakers
Be.Neural - Imperial
Motor Planning Under Uncertainty
Animals use feedback to rapidly correct ongoing movements in the presence of a perturbation. Repeated exposure to a predictable perturbation leads to behavioural adaptation that compensates for its effects.
In this talk, I will primarily focus on a study where we tested the hypothesis that all the computations necessary for motor adaptation may emerge as properties of a controller that adaptively updates its policy. We trained a recurrent neural network to control its own output through an error-based feedback signal, which allowed it to rapidly counteract external perturbations. Implementing a biologically plausible plasticity rule based on this same feedback signal enabled the network to learn to compensate for persistent perturbations through a trial-by-trial process, which approximated gradient descent online. Moreover, the network activity changes during learning matched those from populations of neurons from monkey primary motor cortex —known to mediate both movement correction and motor adaptation— during the same task. Our model also natively reproduced several other behavioural aspects of motor learning in humans and monkeys.
Thus, key features of trial-by-trial motor adaptation can arise from the internal properties of a recurrent neural circuit that adaptively controls its output based on ongoing feedback.
Professor Vijay Mohan Namboodiri
University of California
Implications of Retrospective Causal Learning
Until recently, it was believed that the algorithm and neural circuit for associative learning was understood. The algorithm is temporal difference (TD) reinforcement learning, with the critical teaching signal of a reward prediction error (RPE) conveyed by mesolimbic dopamine. This consensus has recently been challenged by many publications, including ours. My lab developed an alternative theory for associative learning and dopamine function based on the retrospective identification of causes of meaningful outcomes such as rewards. This algorithm makes several surprising and counterintuitive predictions, some of which directly oppose those of the RPE model. My talk will present the results of attempts to test and falsify these predictions.
Princeton Neuroscience Institute
Watching the Watchers: Using quantitative behavior to interrogate social learning during aggression
The ability to observe complex social behavior and use observed information to bias future action is a fundamental building block of social cognition. A foundational unanswered question is whether social observation promotes persistent neural changes that mimic those in individuals that learn through direct experience. We develop a strategy to perform large-scale recordings across subcortical networks for social behavior control and learning and record longitudinally while animals “practice” aggression through direct experience or through observation of aggression. We develop novel methods for behavioral quantification and probe what animals have learned during experience and observation. Using supervised and unsupervised methods for behavioral quantification, we detect unique signatures of a shared behavioral strategy following experience and observation. In addition, we find that observation promotes “experience-like” activity that can recruit a shared plasticity mechanism. These changes bias behavior toward adaptive defensive strategies in new contexts and suggest a similar learned goal emulation strategy.
(Princeton University)
Extended Mean-Field Theories for Networks of Real Neurons
If the behavior of a system with many degrees of freedom can be captured by a small number of collective variables, then plausibly there is an underlying mean-field theory. We show that simple versions of this idea fail to describe the patterns of activity in networks of real neurons. An extended mean-field theory that matches the distribution of collective variables is at least consistent, though shows signs that these networks are poised near a critical point, in agreement with other observations. These results suggest a path to analysis of emerging data on ever larger numbers of neurons.
Columbia
Neural Representational Geometries Reflect Behavioral Differences in Monkeys and Recurrent Neural Networks
Animals likely use a variety of strategies to solve laboratory tasks. Traditionally, combined analysis of behavioral and neural recording data across subjects employing different strategies may obscure important signals and give confusing results. Hence, it is essential to develop techniques that can infer strategy at the single-subject level. We analyzed an experiment in which two male monkeys performed a visually cued rule-based task. The analysis of their performance shows no indication that they used a different strategy. However, when we examined the geometry of stimulus representations in the state space of the neural activities recorded in dorsolateral prefrontal cortex, we found striking differences between the two monkeys. Our purely neural results induced us to reanalyze the behavior. The new analysis showed that the differences in representational geometry are associated with differences in the reaction times, revealing behavioral differences we were unaware of. All these analyses suggest that the monkeys are using different strategies. Finally, using recurrent neural network models trained to perform the same task, we show that these strategies correlate with the amount of training, suggesting a possible explanation for the observed neural and behavioral differences.
Redwood Centre for Theoretical Neuroscience
From McCulloch-Pitts to Retina
Abstract: (Coming soon)
Spotlight Talks
Northeastern University
Transformer Dynamics: A neuroscientific approach to interpretability of large language models
As artificial intelligence models have exploded in scale and capability, understanding of their internal mechanisms remains a critical challenge. Inspired by the success of dynamical systems approaches in neuroscience, here we propose a novel framework for studying computations in deep learning systems. We focus on the residual stream (RS) in transformer models, conceptualizing it as a dynamical system evolving across layers. We find that activations of individual RS units exhibit strong continuity across layers, despite the RS being a non-privileged basis. Activations in the RS accelerate and grow denser over layers, while individual units trace unstable periodic orbits. In reduced-dimensional spaces, the RS follows a curved trajectory with attractor-like dynamics in the lower layers. These insights bridge dynamical systems theory and mechanistic interpretability, establishing a foundation for a "neuroscience of AI" that combines theoretical rigor with large-scale data analysis to advance our understanding of modern neural networks. More details can be found in our pre-print: https://arxiv.org/abs/2502.12131
Imperial College London
Semantic Enrichment of Episodic Memories through Bidirectional Replay in Complementary Learning Systems
Traditional complementary learning systems (CLS) models highlight a predominantly unidirectional information flow from a fast-learning episodic system (medial temporal lobe, MTL) to a slower-learning semantic system (neocortex, CTX). However, episodic memories are not only sensorial, and instead embed a substantial semantic understanding of the episode. Furthermore, MTL regions such as the entorhinal cortex and the hippocampus have been shown to encode semantic representations of the environment, as evidenced by concept cells in humans or splitter cells in rodents. Yet, existing models typically overlook how semantic knowledge could reciprocally influence episodic memory encoding. Here, we introduce a model of cortical-subcortical interactions that addresses these limitations by implementing, on top of bidirectional communication, bidirectional learning. Our framework explicitly models a three-phase learning loop: initial semantic abstraction from sensory episodes, semantic reorganization of MTL (subcortical) representations based on cortical replay, and semantic refinement via replay of sensory-semantic episodes. Crucially, our model distinguishes two different subcortical subpopulations—dense (sensory-based) and sparse (semantic-specific)—facilitating the dynamic restructuring of episodic encodings based on semantic information. Notably, this is achieved using only Hebbian and homeostatic plasticity and no input-target distinction, in contrast with models employing Contrastive Hebbian Learning or backpropagation. The resulting semantic-enriched episodic encodings significantly enhance generalization and compositional capabilities, allowing efficient encoding of never-seen episodes composed of familiar semantic elements. Our model extends classic CLS theories by elucidating the critical role of reciprocal semantic-episodic interactions in fostering flexible and robust memory systems.
Cambridge
A neural network model of continual learning through closed-loop interaction with the environment
Humans continuously acquire, update, and express memories to meet behavioral demands. This continual learning poses a major challenge for artificial neural networks, which often require specialized mechanisms such as memory replay or custom-designed learning rules. Sensorimotor learning offers an ideal test bed for investigating continual learning, as experiments reveal that motor adaptation depends on both recent and distant experiences. While abstract Bayesian models have captured many of these phenomena, the underlying neural principles remain unclear. Here, we demonstrate that several hallmarks of continual motor learning can be explained by a simple principle: continual error-driven learning within a neural network operating in a closed-loop environment. At any moment, the network receives a combination of an efference copy of its output and a supervisory error signal from the previous trial. Synaptic weights are continuously updated using standard gradient-based learning to minimize error. Our approach, supported by experimental data fits, numerical simulations, and analytical derivations, reproduces key features of motor learning. These include savings, where relearning occurs more rapidly than initial learning; the effects of environmental consistency and gradual perturbations on learning; and both spontaneous and evoked recovery, in which memory is expressed even in the absence of the original learning context. Our findings suggest that closed-loop interaction with the environment plays a crucial role in continual learning. By linking behavioral experiments with neural network models, our analytically tractable model provides insights into how error-driven, closed-loop mechanisms can bridge the gap between empirical observations and theories of neural network learning, thus advancing continual learning.
Sainsbury Wellcome Centre UCL
The Prefrontal Cortex is known to play a pivotal role in both value-based decision-making and schema learning. However, the frameworks for each remain largely distinct and there has been little success in unifying the two. We propose that PFC value responses are consistent with the schema framework, by demonstrating that artificial networks learn value representations in order to navigate a cognitive map of the task. Consistent with the influential Miller \& Cohen hypothesis of PFC as an executive controller, these value representations act as control signals which gate the flow of activity. In order to temporally orchestrate gating, value representations dynamically shift within trials, resolving the puzzling observation that PFC value cells are also time-tuned.
Why is there Value in the Prefrontal Cortex?
Estimating Flexible Across-Area Communication With Neurally-Constrained RNNs
Previous work investigating the neural dynamics underlying context-dependent decision making typically analyses a single brain region (typically PFC) or recurrent neural network (RNN). However, evidence suggests that the information required to solve these tasks is distributed across multiple regions. Here, we investigate the neural dynamics across seven brain regions of the non-human primate brain where such distributed information has been observed. By examining within-region geometry and dynamics, we identified significant differences not captured by classical decoding analyses. Using surrogate causal perturbation on multi-regional RNNs trained on condition-averaged data, we explored how inter-area interactions shaped these different neural representations. Our findings reveal that even when task-inputs were withheld from frontal regions during testing, these regions still encoded stimulus information and generated response codes, similar to brain data. Conversely, delivering inputs only to frontal regions or blocking across-region interaction lead to network dynamics that represented stimuli, but failed to solve the task and lacked attractor states for current contexts. Gradually disconnecting regions led to an abrupt breakdown of task-solving capabilities, analogous to spatial bifurcation phenomena. Perturbation experiments highlighted the differential contributions of various regions, offering predictive insights for future experimental validation. These results underscore the critical role of inter-regional communication in task performance and provide a framework for understanding distributed neural processing.