Navigation from first principles
Mate Lengyel · Keynote, Day 1
Navigation is a defining signature of autonomous behaviour that presents animals with fundamental computational challenges. Here we studied how the hippocampal–entorhinal system solves two such challenges. First, efficient navigation requires a continuously updated representation of uncertainty about one's location. We formalised this uncertainty through an image-computable Bayesian ideal observer model that infers location and heading direction as latent variables from self-motion, haptic (when available), and visual sensory inputs, the latter modelled as a retinal image obtained by a pinhole projection. The ideal observer's inferences about location were used to drive homing behaviour in spatial memory tasks and assumed to be represented in the population activity of entorhinal grid cells. Our model accounted for a wide range of ubiquitously described, but puzzling, forms of apparent suboptimalities in navigational behaviour and grid cell responses under deformed environmental geometries. Second, efficient navigation also requires rapid generalisation to novel tasks (goal locations or reward configurations) in a familiar environment. Here we propose the hierarchical successor representation (HSR) by incorporating temporal abstractions into the well-known successor representation (SR). HSR, unlike the classical SR, provides a policy-agnostic multi-scale map that effectively bridges model-free optimality and model-based flexibility, and scales well in topologically complex environments. Furthermore, the HSR successfully accounts for the multi-scale organisation of hippocampal place fields (the distribution of the number of place fields per cell, their sizes, and magnitudes, and how all these depend on the size and structure of the environment). These results suggest that widely described but seemingly idiosyncratic features of neural responses in the hippocampal formation are explained by the first principles of spatial uncertainty and hierarchical behaviour.
Heterogeneity in the brain: lessons from theory and experiment
Fleur Zeldenrust · Session 1
Neurons in the brain are not identical: they show a wide variety of properties. Recently, we showed that this neural diversity itself is not static: it depends on for instance input characteristics and on neuromodulatory state. How does this heterogeneity influence the dynamics and information processing of neural networks? Is this a dial the brain can use to adapt to the requirements of different tasks? I will discuss our latest analysis of experimental data on the heterogeneity of cell properties and how this influences network dynamics and performance in different tasks using different types of network simulations.
Heterogeneous synaptic motifs bridge microscale structure and macroscale nonlinear dynamics
Yuxiu Shao · Session 1
Recent breakthroughs in synaptic-resolution network connectomics have revealed that brain circuits feature fine-scale structural connectivity, such as pairs of correlated synaptic couplings known as second-order motifs. Large-scale recordings of neuronal activity in networks containing nonlinear neurons reveal macroscopic heterogeneous population dynamics throughout cortices. These findings rekindle the inquiry into an intriguing question: may microscale synaptic structures contribute to macroscopic heterogeneous dynamics and computations in ways that canonical brain circuit models cannot? To answer this question, we create random RNNs with various cell types, nonlinear non-negative neural responses, and arbitrary marginal and second-order correlated synaptic statistics. We derive mean-field low-rank equations for P-population networks in which the pre- and postsynaptic neuronal population identities determine the synaptic and motif strengths. Our framework requires 2P latent dynamic variables with P variables describing mean population activity and P variables capturing within-population variability. Theoretical and simulating results demonstrate that chain motifs induce correlations in synaptic variability, enabling microscopic fluctuations to be integrated and influence mesoscopic mean population dynamics. We apply this framework to reverse engineer network connectivity that recapitulates the heterogeneous activity across the population in the mouse primary visual cortex. By bridging the gap between synaptic organisation and nonlinear heterogeneous population dynamics, these findings offer a principled approach and verifiable predictions regarding the relationship between fine-scale connectivity, heterogeneous dynamics, and functional computations.
RNNs perform task computations by dynamically warping neural representations
Arthur Pellegrino · Session 1
Analysing how biological and artificial neural networks represent input features in their activations can help interpret how they perform tasks. In parallel, neuroscience and machine learning have seen a surge of interest in understanding how dynamical systems perform computations on time-varying inputs. Yet, the link between computation-through-dynamics and representational geometry remains poorly understood. In this work, we hypothesise that recurrent neural networks (RNNs) perform computations by dynamically warping their representations of task variables. To test this hypothesis, we develop a Riemannian geometric framework that enables the exact derivation of the neural manifold topology and geometry of a dynamical system from the manifold of its inputs. By characterising the time-varying geometry of RNNs in decision making and working memory tasks, we show that dynamic warping is a fundamental feature of task computation.
Bistable perception reveals a canonical computation in sensory processing
Alex Hyafil · Session 2
Perception is thought to rely on probabilistic inference over a hierarchy of sensory representations. While lateral interactions dominate in the visual cortex, their computational role in perception remains unclear. Here we propose that lateral interactions encode statistical coupling between features represented at the same hierarchical level; that stimulus properties can modulate lateral coupling; and that bistable perception emerges when such coupling becomes sufficiently strong. Within a normative framework of approximate probabilistic inference, strong feature coupling naturally gives rise to a bistable energy landscape, producing hysteresis and bimodal confidence distributions without invoking specialized switching mechanisms. We test these predictions using a structure-from-motion rotating-cylinder stimulus whereby stimulus-induced coupling is manipulated parametrically. As coupling increases, perceptual reports exhibited increased hysteresis, longer dominance durations and over-confidence, as expected. Together, our results suggest that lateral interactions support probabilistic inference by exploiting stimulus statistics, and that bistable perception reflects a regime in which lateral recurrent dynamics dominate perceptual processing.
Input-dependent directionality of interactions between cortical areas
Francesca Mastrogiuseppe · Session 2
Tracking signal flow across brain areas is essential for understanding cognition. Recent studies using cross-covariances show that activity directionality can shift rapidly with behavioral or task demands; yet, the circuit mechanisms underlying these changes remain unclear. Here, we use recurrent network models to investigate how directional interactions emerge and are flexibly reconfigured in multi-area cortical circuits. We show that, for fixed connectivity, directionality is shaped by how common inputs align with recurrent connectivity and the associated internal timescales of activity. In multi-area circuits with locally balanced excitation and inhibition, this reveals a predominant role for inputs to excitatory over inhibitory populations in controlling directionality. These inputs govern the directionality of the latent signals that account for most of the shared activity across areas, predominantly reflecting widespread and coherent activity fluctuations. Finally, our models capture key features of cross-covariances observed in primate V1 and V2 and suggest mechanisms for the shift in directionality reported in these areas.
Single-neuron perturbations reveal a dynamical switch in decision computations in parietal cortex
Giulio Bondanelli · Session 2
Much work has revealed the brain areas and neural activity patterns that contribute to decision-making, leading to long-standing models of how brain circuits perform the underlying computations. However, due to the dense interconnectivity of brain areas and wide-spread representations of decision-related information, an outstanding challenge is to identify how individual brain areas contribute to these computations. To reveal the computations performed by a brain area, we optogenetically stimulate individual neurons as mice perform a decision-making task and measure the resulting activity changes of surrounding neurons. We discover that, as mice make and execute navigation decisions, the posterior parietal cortex (PPC) undergoes a dynamical switch in its computations, from a circuit that helps generate a decision to a circuit that propagates decision information in time as a sequence of neural activity. Early in a trial of a navigation decision task, photostimulating individual choice-selective neurons increases the activity of neighboring neurons that are selective to the same choice and suppresses the activity of neurons preferring the opposite choice. PPC therefore supports decision-making through like-excites-like and opponent inhibition motifs. In contrast, later in the trial during a delay period between visual cues and decision reports, these motifs are absent, and instead photostimulating choice-selective neurons excites neurons later in a sequence of neural activity and inhibits those earlier in the sequence. Thus, the same anatomical circuit can rapidly switch to distinct computational regimes without changes in connectivity. These results reveal a causal role for PPC in navigational decision-making, both in decision formation and decision maintenance.
TBA
Cristina Savin · Keynote, Day 1
Abstract to follow.
Mechanisms for multi-tasking in recurrent neural networks
Ashok Litwin-Kumar · Keynote, Day 2
The ability to generate different responses across changing goals and circumstances is a defining feature of intelligent behavior. How such multitasking can be achieved, without interference across conditions, remains an open question. We describe and compare strategies a single recurrent network may employ to achieve multi-tasking without one task interfering with another. Using the theory of low-rank recurrent neural networks, we demonstrate that previous models of multi-tasking can be organized based on their mechanism for switching between tasks and avoiding interference. We introduce a new model of "task barcodes" that enable flexible input-driven task selection beyond what is achieved by previous theoretical models. We analyze concrete implementations of these mechanisms and show that each produces qualitatively different features in the dynamical system governing the evolution of neural firing rates.
Motor control modelling
Heike Stein · Session 3
Topic: dynamical systems reconstruction of behavioral data, some biomechanics and optimal control perspectives, and possibly a bit of cerebellar data.
Preparatory population dynamics of timing and flexible behavior
Lorenzo Fontolan · Session 3
Acting at the right moment is a central component of flexible behavior. I previously showed that preparatory activity in frontal cortex achieves robustness via attractor dynamics, i.e. stable population patterns, and that temporal modulation of such attractors determines the circuit's sensitivity to incoming signals. How such temporal modulation is implemented in neural circuits, however, remains unknown. Combining perturbations, multi-regional electrophysiology, and computational modeling, we found that flexible motor timing appears to be governed by integrator dynamics distributed across the cortico-basal ganglia loop, in which the striatum integrates cortical input to generate a graded timing signal that determines when an action is initiated. Using data-constrained network models, we identified the striatum as the locus of integration, indicating distinct dynamical roles for the two nodes despite similar ramping activity in both regions. Our results suggest that the cortico-striatal integrator provides the temporal drive that modulates attractor dynamics in frontal cortex, linking the stability of motor preparation to the flexible control of action release. This view of preparation as a temporally modulated population state also raises a dimensional question that we addressed in separate work: which components of preparatory activity influence future behavior at different moments in time? We found that these behaviorally relevant dimensions can be low-amplitude and distinct from the dominant trajectories of neural activity. Thus, flexible behavior depends not only on the low-dimensional organization of preparatory dynamics, but also on how behavioral influence is distributed across dimensions over time.
Primary and Supplementary Motor cortex implement parallel solutions for the control of rhythmic and discrete arm movements
Mark Humphries · Session 3
Arm movements are rhythmic, discrete, or some combination of the two. Conflicting evidence supports each of two possible solutions for how motor cortex controls them: that either it uses the same strategy for controlling rhythmic and discrete movements or different strategies for distinct movement types. Using recurrent neural network modelling and multi-unit recordings during an arm-cycling task, we show that primate motor cortex uses both solutions. Primary motor cortex (M1) dynamics converge to the same limit-cycle when executing both movement types. In contrast, supplementary motor area (SMA) dynamics diverge according to the type of the upcoming movement before reaching a helical spiral. Our results reconcile opposing views on the cortical control of rhythmic and discrete movements by showing that the two solutions are not mutually exclusive but implemented in parallel within motor cortex. With Andrea Colins Rodriguez & Romulo Fuentes.
Computational Modeling of Rigid Beliefs: An illustration with Hallucinations, Delusions and Conspiracy
Renaud Jardri · Session 4
In this lecture, I will explore the computational mechanisms underlying rigid beliefs, ranging from clinical hallucinations to the widespread adherence to conspiracy theories. By moving beyond standard Bayesian cognition, which views the brain as a predictive machine balancing prior knowledge with sensory evidence, I will present the Circular Inference Model, developed with Sophie Denève during my time at GNT. This framework suggests that a breakdown in inhibitory control (specifically within hierarchical cortical loops) causes the brain to "double-count" information, leading to overconfidence and distorted perceptions. I will present experimental data from the "Fisher Task" as well as from bistable perception tests, which demonstrate that both patients with schizophrenia, but also individuals with high conspiracy ideations, exhibit non-linear deviations from Bayesian optimality. Interestingly, longitudinal data collected during major political events tends to suggest that after an initial sensory amplification through these circular loops, strengthening prior beliefs may serve as a psychological coping mechanism. Doing so, individuals could reduce their subjective stress and regain a sense of stability in the face of high environmental uncertainty at the cost of suboptimal reasoning.
Dendritic contribution to learning efficient codes
Viola Priesemann · Session 4
We propose an unsupervised predictive learning framework based on dendritic error computation, incorporating heterosynaptic excitatory and inhibitory plasticity. A central feature is adaptive dendritic inhibition that locally balances depolarization, thereby gating voltage-dependent excitatory plasticity. This ensures that synaptic changes are driven by prediction errors rather than raw inputs, consistent with experimental findings across multiple brain systems, and consistent with prior efficient coding theories. We show that this mechanism allows neurons to learn non-redundant representations of complex spatiotemporal inputs without supervision, highlighting the putative contribution of any dendritic branch to learning efficient codes.
Machine learning for mechanistic insight in neuroscience
Pedro Goncalves · Session 4
Mechanistic modeling aims to explain observed phenomena in terms of underlying causes, but identifying the parameters for which a model is consistent with complex and stochastic data is challenging. I will describe our recent developments in simulation-based inference allowing us to scalably retrieve the model parameters from data, and how we apply them to a range of neuroscience problems at the single neuron and neural circuit levels.
Characterizing the dynamics of learning from animal training data
Jonathan Pillow · Keynote, Day 2
The dynamics of learning in natural and artificial environments is a problem of great interest to both neuroscientists and artificial intelligence experts. In this talk, I will describe recent studies from my lab that focus on characterizing the internal state dynamics governing sensory decision making, both in fully trained mice and in mice learning to perform a new task. First, I will describe work showing that even expert mice alternate between multiple strategies, characterized by distinct states of a hidden Markov model (HMM), when performing a sensory decision-making task. I will then describe a dynamic extension of this model to show that these states emerge early during training. Finally, I will describe recent work on inferring learning rules from behavioral data. This work combines ideas from reinforcement learning and adaptive experimental design to infer animal learning rules and use these rules to speed up animal training.