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 both considering mechanistic neural model 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)
Dr Francesca Mastrogiuseppe (Champalimaud Center for the Unknown)
Keynote Talks
Professor Peter Dayan (Max Planck Institute, Tübingen): Controlling the Controller: Instrumental Manipulations of Pavlovian Influences via Dopamine
Professor Mackenzie Mathis (EPFL): Learnable Neural Dynamics
Invited Talks
Professor Athena Akrami (UCL): Circuits and computations for learning and exploiting sensory statistics
Professor Nicolas Brunel (Duke): Roles of inhibition in shaping the response of cortical networks
Dr Sophia Sanborn (Science): Symmetry and Universality
Dr Lea Duncker (Stanford): Evaluating dynamical systems hypotheses using direct neural perturbations
Dr Kris Jensen (UCL): An attractor model of planning in frontal cortex
Spotlight Talks
Cristiano Capone (ISS): Online network reconfiguration: non-synaptic learning in RNNs
Sam Hall-McMaster (Harvard University): Neural Prioritisation of Past Solutions Supports Generalisation
Alexander Mathis (EPFL): Modeling sensorimotor circuits with machine learning: hypotheses, inductive biases, latent noise and curricula
Stefano Diomedi (NRC Italy): Neural subspaces in three Parietal areas during reaching planning and execution
Sofia Raglio (Sapienza): Clones of biological agents solving cognitive task: hints on brain computation paradigms
Arno Granier (Bern): Confidence estimation and second-order errors in cortical circuits
Erik Hermansen (NTNU): The Ontogeny of the Grid Cell Network – Uncovering the Topology of Neural Representations
Steeve Laquitaine (EPFL): Cell types and layers differently shape the geometry of neural representations in a biophysically detailed model of the neocortical microcircuit.
Subhadra Mokashe (Brandeis University): Competition between memories for reactivation as a mechanism for long-delay credit assignment
Brendan A. Bicknell (UCL): Fast and slow synaptic plasticity enables concurrent control and learning
Vezha Boboeva (Sainsbury Wellcome Centre, UCL): Computational principles underlying the learning of sequential regularities in recurrent networks
Keynote Talks
Max Planck Institute, Tübingen
Controlling the Controller: Instrumental Manipulations of Pavlovian Influences via Dopamine
Pavlovian influences notoriously interfere with operant behaviour - with dopamine being one potential culprit. Here, using the examples of active avoidance and omission behaviour, we examine the possibility that direct manipulation of the dopamine signal is an instrument of control itself. We argue that dopamine levels might be affected by the controlled deployment of a reframing mechanism that recasts the prospect of possible punishment as an opportunity to approach safety, and the prospect of future reward in terms of a possible loss of that reward. We model two canonical experiments, showing that we can capture critical features of both behaviour and dopamine transients. This is joint work with Kevin Lloyd, Azadeh Nazemorroaya and Dan Bang.
EPFL
Learnable Neural Dynamics
Mapping behavioral actions to neural activity is a fundamental goal in neuroscience. One emerging way to study neural circuits is through the lens of neural dynamics with machine learning. In this talk, I will introduce a new method, CEBRA, which utilizes both behavioral and neural data either through hypothesis-driven or discovery-driven approaches, producing consistent, high-performance lower dimensional representations of neural dynamics. I will also discuss ongoing work at providing identifiable attribution of individual neurons to these latent dynamics, which could pave the way for more interpretability.
Invited Talks
University College London (UCL)
Circuits and computations for learning and exploiting sensory statistics
A defining feature of animal intelligence is the ability to discover and update knowledge of statistical regularities in the sensory environment, in service of adaptive behaviour. This allows animals to build appropriate priors, in order to disambiguate noisy inputs, make predictions and act more efficiently. Despite decades of research in the field of human cognition and theoretical neuroscience, it is not known how such learning can be implemented in the brain. By combing highly quantifiable cognitive tasks in humans, rats, and mice, as well as neuronal measurements and perturbations in the rodent brain and computational modelling, we seek to build a multi-level description of how sensory history is utilised in inferring regularities in temporally extended tasks. In this talk, I will specifically focus on a cross-species model to study learning and exploiting statistical prior distributions in working memory and sensory discrimination behaviours.
Duke University
Roles of inhibition in shaping the response of cortical networks
Normalization is a key function of sensory cortices, allowing detection of weak stimuli without overexcitation to strong stimuli. Normalization naturally occurs in inhibition stabilized networks (ISNs), in which recurrent inhibition balances rising recurrent excitation. While there is evidence that sensory cortex is inhibition-stabilized, how and when different types of interneurons contribute to inhibition stabilization is poorly understood. Somatostatin+ interneurons (SST) are strongly recurrently connected with neighboring Pyramidal cells (Pyr) and contribute to normalization and shaping Pyr output in high arousal states, but their specific role in ISN dynamics remains unclear. To better understand the role of SST cells in network stabilization, we combined cell type-specific pharmacology with 2-photon calcium imaging
to record activity of SST cells and neighboring Pyr cells in mouse primary visual cortex (V1) before and after antagonizing AMPA receptors on SST cells. This antagonism suppressed SST responses to weak visual stimuli, but the suppressive effect was attenuated or by strong stimuli o locomotion, and inverted by both together. Analysis of a computational model including Pyr, SST and Parvalbumin+ (PV) cells revealed that this data is consistent with a network that is stabilized purely by PV cells in stationary conditions and with weak stimuli, but
that requires SST cells for stabilization during locomotion with strong sensory stimuli. These results elucidate the conditions under which SST cells are necessary to stabilize cortical circuits.
Science
Symmetry and Universality
Artificial neural networks trained on natural data exhibit a striking phenomenon: regardless of exact initialization, dataset, or training objective, models trained on the same data domain frequently converge to similar learned representations — a phenomenon known as convergent learning. The first layers of diverse image models, for example, tend to learn Gabor filters and color-contrast detectors. Remarkably, many of these same features are observed in the visual cortex, suggesting the existence of “universal” representations that transcend biological and artificial substrate. In this talk, I will present theoretical work that explains the phenomenon of convergent learning as a byproduct of the symmetries of natural data—i.e. the transformations that leave perceptual content invariant. We provide a mathematical proof that certain features (i.e. harmonics) are guaranteed to emerge in neural networks trained on tasks that require invariance to the actions of a group of transformations, such as translation or rotation. Given the centrality of invariance to many perceptual tasks, our proof is able to explain a broad class of convergent learning phenomena in a wide class of neural network architectures, while providing a new perspective on classical neuroscience results. This work sets a foundation for understanding universal principles of neural representation with the mathematics of symmetry.
Joint work with Giovanni Luca Di Marchetti, Christopher Hillar, & Danica Kragic
Stanford University
Evaluating dynamical systems hypotheses using direct neural perturbations
The rich repertoire of skilled mammalian behavior is the product of neural circuits that generate robust and flexible patterns of activity distributed across populations of neurons. Decades of associative studies have linked many behaviors to specific patterns of population activity, but association alone cannot reveal the dynamical mechanisms that shape those patterns. Are local neural circuits high-dimensional dynamical reservoirs able to generate arbitrary superpositions of patterns with appropriate excitation? Or might circuit dynamics be shaped in response to behavioral context so as to generate only the low-dimensional patterns needed for the task at hand? Here, we address these questions within primate motor cortex by delivering optogenetic and electrical microstimulation perturbations during reaching behavior. Our results demonstrate that motor cortical activity during reaching is shaped by a self-contained, low-dimensional dynamical system. The subspace containing task-relevant dynamics proves to be oriented so as to be robust to strong non-normal amplification within cortical circuits. This task dynamics space exhibits a privileged causal relationship with behavior, in that stimulation in motor cortex perturb reach kinematics only to the extent that it alters neural states within this subspace. Our results resolve long-standing questions about the dynamical structure of cortical activity associated with movement, and illuminate the dynamical perturbation experiments needed to understand how neural circuits throughout the brain generate complex behavior.
University College London (UCL)
An attractor model of planning in frontal cortex
Animals can flexibly navigate complex and changing environments. Prior work suggests that this flexibility is facilitated by an internal world model, which supports planning and decision making. Prefrontal cortex has commonly been implicated as an important region for such model-based planning, yet little is known about the underlying algorithms and neural implementations. We suggest a putative circuit model of planning in prefrontal cortex, inspired by recent experimental findings of conjunctive representations of space and time. In this model, a recurrent neural network receives start and goal states as inputs. The fixed points of the dynamics are the paths between these terminal states, with each intermediate state represented in a separate neural population. The recurrent network dynamics transform an initial global activity state to such a path, which is updated online as the agent acts out the plan. This model also generalizes to hierarchical planning and goals that change predictably over time, which are important challenges for biological organisms in naturalistic environments.
Spotlight talks
Istituto Superiore di Sanità, Italy
Online network reconfiguration: non-synaptic learning in RNNs
Behavioral adaptation in humans, as a consequence of an error or a verbal instruction, can be extremely rapid. Improvement in performances are usually associated in machine learning to optimization of network parameters such as synaptic weights. However, such rapid changes are not coherent with the timescales of synaptic plasticity, suggesting that the mechanism responsible for that could be a dynamical network reconfiguration. Similar capabilities have been observed in transformers, foundational architectures in the field of machine learning that are widely used in applications such as natural language and image processing. Transformers are capable of in-context learning, the ability to adapt and acquire new information dynamically within the context of the task, without the need for changes to their underlying parameters. We propose that a similar mechanism can be introduced in a recurrent neural network by considering a temporal dynamics on an attention mechanism that changes the way input is integrated, converging to the proper solution without synaptic plasticity. We argue that such a framework reproduces the psychometry of context-dependent tasks in humans, solving the incoherence of plasticity timescales.
Harvard University
Neural Prioritisation of Past Solutions Supports Generalisation
How do we decide what to do in new situations? One way to solve this dilemma is to reuse solutions developed for other situations. There is now some evidence that a computational process capturing this idea – called successor features & generalised policy improvement – can account for how humans transfer prior solutions to new situations. Here we asked whether a simple formulation of this idea could explain human brain activity in response to new tasks. Participants completed a multi-task learning experiment during fMRI (n=40). The experiment included training tasks that participants could use to learn about their environment, and test tasks to probe their generalisation strategy. Behavioural results showed that people learned optimal solutions (policies) to the training tasks, and reused them on test tasks in a reward-selective manner. Neural results showed that optimal solutions from the training tasks received prioritised processing during test tasks in occipitotemporal cortex and dorsolateral prefrontal cortex. These findings suggest that humans evaluate and generalise successful past solutions when solving new tasks.
EPFL
Modeling sensorimotor circuits with machine learning: hypotheses, inductive biases, latent noise and curricula
Hierarchical sensorimotor processing, modularity and experience are all essential for adaptive motor control. Recent efficient musculoskeletal simulators and machine learning algorithms provide new computational approaches to gain insights into those concepts for biological motor control. Firstly, I will present a hypothesis-driven modeling framework to quantitatively assess the computations underlying proprioception. We trained thousands of models to transform muscle spindle inputs according to 16 hypotheses from the literature. For all those hypotheses, we found that hierarchical models that better satisfy those hypotheses, also explain neural recordings in the brain stem and cortex better. We furthermore find that models trained to estimate the state of the body are best at explaining neural data. Secondly, I will discuss methods to close the gap between reinforcement learning algorithms and biological motor control. I will highlight several ingredients (brain-inspired inductive biases, latent noise, curriculum learning) for learning controllers with high-dimensional musculoskeletal systems. Taken together, these results highlight the importance of inductive biases, and experience for biological motor control.
Institute of Cognitive Sciences and Technologies (ISTC), National Research Council of Italy (CNR), Padua, Italy
Neural subspaces in three Parietal areas during reaching planning and execution
The posterior parietal cortex (PPC) plays a crucial role in planning and executing reaching movements. Earlier, we showed the difficulty of identifying distinct subpopulations of neurons based on their functions, due to mixed selectivity. We also found a sequence of neural states that align with the different motor stages of the reaching movement, which rely on whole populations activity. Thus, the same neurons can perform different computations based on the movement phase, raising questions about how the entire population is organised to allow such a flexibility. To investigate this, here we characterize the neural subspaces in three PPC areas (V6A, PEc, PE) during a reaching task. We applied Principal Component Analysis via manifold optimization and identified orthogonal subspaces for movement planning and execution in the somatomotor area PE, finding independent neural dynamics in the two motor stages. To the contrary, shared subspaces between the two epochs can be identified in the visuomotor areas V6A and PEc. The results demonstrate the existence of different population-level subspaces across the three parietal areas examined, enabling the PPC to perform different epoch-specific computations using the same neural substrate.
La Sapienza University of Rome
Clones of biological agents solving cognitive task: hints on brain computation paradigms
Understanding brain computation is a significant challenge in neuroscience, often revealing different paradigms compared to those of Machine Learning. Our study introduces a novel approach, using electrophysiological data, to create 'digital clones' - biologically plausible computational models performing the same cognitive tasks. Unlike traditional AI architectures, our model incorporates sophisticated interaction dynamics between mean-field populations, capturing the intrinsic computational logic of the brain. Here we focus on transitive inference (TI) task, where participants are presented with adjacent pairs from a series of arbitrarily ordered items, and asked to infer the relative order of items never shown together during the training. We recorded the multi-unit activity (MUA) from the premotor cortex of rhesus monkeys performing TI task. We define a rate model reproducing simultaneously the MUA from experiments, and the behavioral output of the agent. Remarkably, despite the linearity of our model, it reproduces the most important aspects of observed data. Moreover, the comparison of parameters inferred in different phases of the task might provide information on how brain processing evolves during learning.
University of Bern
Confidence estimation and second-order errors in cortical circuits
Minimization of cortical prediction errors has been considered a key computational goal of the cerebral cortex underlying perception, action and learning. However, it is still unclear how the cortex should form and use information about uncertainty in this process. Here, we formally derive neural dynamics of predictive coding under the assumption that cortical areas must not only predict the activity in other areas and sensory streams but also jointly estimate their confidence (inverse expected uncertainty) in their predictions. In the resulting neuronal dynamics, the integration of bottom-up and top-down cortical streams is dynamically modulated based on confidence, in accordance with the Bayesian principle. Moreover, the theory predicts the existence of cortical second-order errors, comparing confidence and actual performance. Second-order errors are also propagated through the cortical hierarchy, leading to qualitative enhancements of classification capabilities in single areas, and are used to learn the weights of synapses responsible for estimating confidence. We propose a detailed mapping of the theory to cortical circuitry, discuss entailed functional interpretations and provide potential directions for experimental work.
Norwegian University of Science and Technology
The Ontogeny of the Grid Cell Network – Uncovering the Topology of Neural Representations
Groups of neurons collectively perform computations, their joint activity representing internal or external covariates. Identifying these groups and their representations has often been done with a priori knowledge of the relevant covariate, recording the variable alongside the neural data. However, this has for instance prohibited studying the formation of neural ensembles during development - before the stimuli is available to the animal. While artificial neural networks are usually randomly or uniformly initialized, such a tabula rasa may be inefficient for the networks of the brain. For instance, all mammals need to accomplish many of the same tasks - such as navigate in an (often) two-dimensional world - and may benefit from having preconfigured structures which can aptly grapple with these challenges. Grid cells are thought to play a critical role in this task, selectively active in hexagonally arranged fields in an environment. Using topological data analysis, we have previously shown that the population activity of such cells form a toroidal state space, regardless of the behavioral state. Expanding on this framework, we analysed the activity patterns of correlated clusters of neurons, identifying spatially tuned ensembles such as grid cell modules in head-fixed mice in an unsupervised fashion. More recently, this approach has allowed us to detect neural ensembles with ring-like and toroidal structures in Neuropixels recordings from the entorhinal cortex in rats, from as early as postnatal day 10 - before spatial exploration and opening of eyes and ear-canals. This suggests head direction cells and grid cells may organize independent of experience, and shows we can uncover neural representations solely from the recordings, allowing us to better understand the intrinsic mechanisms of the brain.
EPFL
Cell types and layers differently shape the geometry of neural representations in a biophysically detailed model of the neocortical microcircuit
A central problem in Neuroscience is understanding how the neocortical microcircuit encodes stimuli for perceptual decisions. Theoretical models of cortical networks are mathematically convenient tools that have produced deep insights into cortical computation principles. Yet, their lack of realism does not allow to fully probe the role of the cortex’ biological diversity in shaping sensory representations for perceptual discrimination. We addressed these questions using an extensively validated biophysically detailed model of the rat’s S1 neocortical column. The model comprises 30,190 morphologically detailed neurons, spanning all six cortical layers and receiving inputs from simulated thalamic fibers. It captures the biological diversity in 60 morphological and 11 electrical neuron types and features realistic synaptic connectivity and short-term plasticity. We simulated the microcircuit’s responses to whisker deflections in 360 orientations, by injecting currents into distinct groups of thalamic fibers. We then linked individual neuron’s orientation tuning to the geometry and discrimination capacity of the evoked neural manifold. We found that neurons contribute differently to discrimination capacity based on type and layer.
Brandeis University
Competition between memories for reactivation as a mechanism for long-delay credit assignment
Animals learn to associate an event with its outcome, as in conditioned taste aversion when they gain aversion to a conditioned stimulus (CS, recently experienced taste) if sickness is later induced. If there is another intervening taste (interfering stimulus, IS), the IS gains some credit for the causality of the outcome, reducing aversion to the CS. The known short-term correlational plasticity mechanisms do not wholly explain how networks of neurons achieve long-delay credit assignment. We hypothesize that reactivation of prior events at the time of outcome causes specific associative learning between those events and the outcome. We explore the credit assignment using a spiking neural network model storing memories of two events that inherently compete to be the cause. As one cause becomes more likely, the other becomes less likely to be the cause of the outcome. We explore parameters that influence the degree of competition between the two memories for reactivation. We show how a later memory can be reactivated more often and reduce the reactivation of a prior memory. By reactivating the memories in a probabilistic way, neural networks could perform Bayesian inference to assign the credit in a biologically plausible way.
University College London
Fast and slow synaptic plasticity enables concurrent control and learning
Natural intelligence is rooted in the ability to adapt on multiple timescales. Indeed, this is critical for survival; animals must be able to form lifelong memories, but also react rapidly to disturbances and maintain stable brain activity. Much of this is driven by synaptic plasticity, which exhibits a comparable range of dynamics. To understand the brain, it is therefore imperative to identify its many interacting plasticity rules. Towards this goal, here we develop a normative theory of synaptic plasticity that explains how the output of a neuron can be optimized through concurrent fast and slow mechanisms. We consider a general task in which a neuron must modify its synapses in order to drive a downstream process to match a time-varying target. By framing synaptic plasticity as a stochastic control problem, we derive a biologically plausible update rule that dramatically outperforms classical gradient-based approaches. In this, fast synaptic weight changes greedily correct downstream errors, while slow synaptic weight changes implement statistically optimal learning. Applied in a cerebellar microcircuit, the theory explains widely observed features of spiking behavior and plasticity, and makes novel experimental predictions.
Sainsbury Wellcome Centre, UCL
Computational principles underlying the learning of sequential regularities in recurrent networks
The world, despite its complexity, harbors patterns and regularities crucial for animals, with numerous real-life processes evolving over time into structured sequences of events. Brains have evolved to learn and exploit these sequential regularities, by forming knowledge at different degrees of abstraction: from simple transition and timing, to chunking, ordinal knowledge, algebraic patterns, and finally nested tree structures. Moreover, fragmentary evidence suggests that the most complex classes, specifically algebraic patterns and nested tree structures, may be uniquely human faculties, underpinning distinct human capabilities in language, music, and mathematics. Whether and how certain regularities expressed in "algebraic patterns" or abstract schemas (e.g., AAB) are encoded in the brain is still an open question. Here, we study whether and how neural circuits may acquire, organize and use such an abstract code. We first build a computational framework to generate sequences of different complexity, and design appropriate quantitative measures of "regularity". Next, we propose RNN models capable of performing different tasks requiring learning and predicting such sequences, and study the conditions under which learning is possible, such as the effect of curricula. We study the internal representations formed by the network, and the extent to which these might be abstract, allowing to generalize to novel, unseen sequences. Finally, we study whether and how such representations may be used compositionally, generalizing across different tasks.