van Vreeswijk TNS

The dominant theoretical framework to account for reinforcement learning in the brain is temporal difference (TD) reinforcement learning. The TD framework predicts that some neuronal elements should represent the reward prediction error (RPE), which means they signal the difference between the expected future rewards and the actual rewards. The prominence of the TD theory arises from the observation that firing properties of dopaminergic neurons in the ventral tegmental area appear similar to those of RPE model-neurons in TD learning. Previous implementations of TD learning assume a fixed temporal basis for each stimulus that might eventually predict a reward. Here we show that such a fixed temporal basis is implausible and that certain predictions of TD learning are inconsistent with experiments. We propose instead an alternative theoretical framework, coined FLEX (Flexibly Learned Errors in Expected Reward). In FLEX, feature specific representations of time are learned, allowing for neural representations of stimuli to adjust their timing and relation to rewards in an online manner. In FLEX dopamine acts as an instructive signal which helps build temporal models of the environment. FLEX is a general theoretical framework that has many possible biophysical implementations. In order to show that FLEX is a feasible approach, we present a specific biophysically plausible model which implements the principles of FLEX. We show that this implementation can account for various reinforcement learning paradigms, and that its results and predictions are consistent with a preponderance of both existing and reanalyzed experimental data.

When interacting with complex environments, humans can rapidly adapt their behavior to changes in task or context. To facilitate this adaptation, we often spend substantial periods of time contemplating possible futures before acting. For such planning to be rational, the benefits of planning to future behavior must at least compensate for the time spent thinking. Here we capture these features of human behavior by developing a neural network model where not only actions, but also planning, are controlled by prefrontal cortex. This model consists of a meta-reinforcement learning agent augmented with the ability to plan by sampling imagined action sequences drawn from its own policy, which we refer to as 'rollouts'. Our results demonstrate that this agent learns to plan when planning is beneficial, explaining the empirical variability in human thinking times. Additionally, the patterns of policy rollouts employed by the artificial agent closely resemble patterns of rodent hippocampal replays recently recorded in a spatial navigation task, in terms of both their spatial statistics and their relationship to subsequent behavior. Our work provides a new theory of how the brain could implement planning through prefrontal-hippocampal interactions, where hippocampal replays are triggered by - and in turn adaptively affect - prefrontal dynamics.

Running, swimming, or flying through the world, animals are constantly making decisions while on the move—decisions that allow them to choose where to eat, where to hide, and with whom to associate. Despite this most studies have considered only on the outcome of, and time taken to make, decisions. Motion is, however, crucial in terms of how space is represented by organisms during spatial decision-making. Employing a range of new technologies, including automated tracking, computational reconstruction of sensory information, and immersive ‘holographic’ virtual reality (VR) for animals, experiments with fruit flies, locusts and zebrafish (representing aerial, terrestrial and aquatic locomotion, respectively), I will demonstrate that this time-varying representation results in the emergence of new and fundamental geometric principles that considerably impact decision-making. Specifically, we find that the brain spontaneously reduces multi-choice decisions into a series of abrupt (‘critical’) binary decisions in space-time, a process that repeats until only one option—the one ultimately selected by the individual—remains. Due to the critical nature of these transitions (and the corresponding increase in ‘susceptibility’) even noisy brains are extremely sensitive to very small differences between remaining options (e.g., a very small difference in neuronal activity being in “favor” of one option) near these locations in space-time. This mechanism facilitates highly effective decision-making, and is shown to be robust both to the number of options available, and to context, such as whether options are static (e.g. refuges) or mobile (e.g. other animals). In addition, we find evidence that the same geometric principles of decision-making occur across scales of biological organisation, from neural dynamics to animal collectives, suggesting they are fundamental features of spatiotemporal computation.

Theories from reinforcement learning have been highly influential for interpreting neural activity in the biological circuits critical for animal and human learning. Central among these is the identification of phasic activity in dopamine neurons as a reward prediction error signal that drives learning in basal ganglia and prefrontal circuits. However, recent findings suggest that dopaminergic prediction error signals have access to complex, structured reward predictions and are sensitive to more properties of outcomes than learning theories with simple scalar value predictions might suggest. Here, I will present recent work in which we probed the identity-specific structure of reward prediction errors in an odor-guided choice task and found evidence for multiple predictive “threads” that segregate reward predictions, and reward prediction errors, according to the specific sensory features of anticipated outcomes. Our results point to an expanded class of neural reinforcement learning algorithms in which biological agents learn rich associative structure from their environment and leverage it to build reward predictions that include information about the specific, and perhaps idiosyncratic, features of available outcomes, using these to guide behavior in even quite simple reward learning tasks.

The mapping from the visual field to V1 can be approximated by a log-polar transform. In this domain, scale is a left-right shift, and rotation is an up-down shift. When fed into a standard shift-invariant convolutional network, this provides scale and rotation invariance. However, translation invariance is lost. In our model, this is compensated for by multiple fixations on an object. Due to the high concentration of cones in the fovea with the dropoff of resolution in the periphery, fully 10 degrees of visual angle take up about half of V1, with the remaining 170 degrees (or so) taking up the other half. This layout provides the basis for the central and peripheral pathways. Simulations with this model closely match human performance in scene classification, and competition between the pathways leads to the peripheral pathway being used for this task. Remarkably, in spite of the property of rotation invariance, this model can explain the inverted face effect. We suggest that the standard method of using image coordinates is the wrong prior for models of primate vision.

I will discuss work with Jack Lindsey modeling reinforcement learning for action selection in the basal ganglia. I will argue that the presence of multiple brain regions, in addition to the basal ganglia, that contribute to motor control motivates the need for an off-policy basal ganglia learning algorithm. I will then describe a biological implementation of such an algorithm that predicts tuning of dopamine neurons to a quantity we call "action surprise," in addition to reward prediction error. In the same model, an implementation of learning from a motor efference copy also predicts a novel solution to the problem of multiplexing feedforward and efference-related striatal activity. The solution exploits the difference between D1 and D2-expressing medium spiny neurons and leads to predictions about striatal dynamics.

The cell assembly hypothesis postulates that groups of coordinated neurons form the basis of information processing. Here, we test this hypothesis by analyzing massively parallel spiking activity recorded in monkey motor cortex during a reach-to-grasp experiment for the presence of significant ms-precise spatio-temporal spike patterns (STPs). For this purpose, the parallel spike trains were analyzed for STPs by the SPADE method (Stella et al, 2019, Biosystems), which detects, counts and evaluates spike patterns for their significance by the use of surrogates (Stella et al, 2022 eNeuro). As a result we find STPs in 19/20 data sets (each of 15min) from two monkeys, but only a small fraction of the recorded neurons are involved in STPs. To consider the different behavioral states during the task, we analyzed the data in a quasi time-resolved analysis by dividing the data into behaviorally relevant time epochs. The STPs that occur in the various epochs are specific to behavioral context - in terms of neurons involved and temporal lags between the spikes of the STP. Furthermore we find, that the STPs often share individual neurons across epochs. Since we interprete the occurrence of a particular STP as the signature of a particular active cell assembly, our interpretation is that the neurons multiplex their cell assembly membership. In a related study, we model these findings by networks with embedded synfire chains (Kleinjohann et al, 2022, bioRxiv 2022.08.02.502431).

Learning in neural networks requires assigning the right values to thousands to trillions or more of individual connections, so that the network as a whole produces the desired behavior. Neuroscientists have gained insights into this “credit assignment” problem through decades of experimental, modeling, and theoretical studies. This has suggested key roles for synaptic eligibility traces and top-down feedback signals, among other factors. Here we study the potential contribution of another type of signaling that is being revealed in greater and greater fidelity by ongoing molecular and genomics studies. This is the set of modulatory pathways local to a given circuit, which form an intriguing second type of connectome overlayed on top of synaptic connectivity. We will share ongoing modeling and theoretical work that explores the possible roles of this local modulatory connectome in network learning.

Computation in neural circuits relies on a common set of motifs, including divergence of common inputs to parallel pathways, convergence of multiple inputs to a single neuron, and nonlinearities that select some signals over others. Convergence and circuit nonlinearities, considered individually, can lead to a loss of information about the inputs. Past work has detailed how to optimize nonlinearities and circuit weights to maximize information, but we show that selective nonlinearities, acting together with divergent and convergent circuit structure, can improve information transmission over a purely linear circuit despite the suboptimality of these components individually. These nonlinearities recode the inputs in a manner that preserves the variance among converged inputs. Our results suggest that neural circuits may be doing better than expected without finely tuned weights.

Modern electrophysiological recordings simultaneously capture single-unit spiking activities of hundreds of neurons. These neurons exhibit highly complex coordination patterns. Where does this complexity stem from? One candidate is the ubiquitous heterogeneity in connectivity of local neural circuits. Studying neural network dynamics in the linearized regime and using tools from statistical field theory of disordered systems, we derive relations between structure and dynamics that are readily applicable to subsampled recordings of neural circuits: Measuring the statistics of pairwise covariances allows us to infer statistical properties of the underlying connectivity. Applying our results to spontaneous activity of macaque motor cortex, we find that the underlying network operates in a strongly recurrent regime. In this regime, network connectivity is highly heterogeneous, as quantified by a large radius of bulk connectivity eigenvalues. Being close to the point of linear instability, this dynamical regime predicts a rich correlation structure, a large dynamical repertoire, long-range interaction patterns, relatively low dimensionality and a sensitive control of neuronal coordination. These predictions are verified in analyses of spontaneous activity of macaque motor cortex and mouse visual cortex. Finally, we show that even microscopic features of connectivity, such as connection motifs, systematically scale up to determine the global organization of activity in neural circuits.

Neurons in the rodent hippocampus appear to encode the position of the animal in physical space during movement. Individual ``place cells'' fire in restricted sub-regions of an environment, a feature often taken as evidence that the hippocampus encodes a map of space that subserves navigation. But these same neurons exhibit complex responses to many other variables that defy explanation by position alone, and the hippocampus is known to be more broadly critical for memory formation. Here we elaborate and test a theory of hippocampal coding which produces place cells as a general consequence of efficient memory coding. We constructed neural networks that actively exploit the correlations between memories in order to learn compressed representations of experience. Place cells readily emerged in the trained model, due to the correlations in sensory input between experiences at nearby locations. Notably, these properties were highly sensitive to the compressibility of the sensory environment, with place field size and population coding level in dynamic opposition to optimally encode the correlations between experiences. The effects of learning were also strongly biphasic: nearby locations are represented more similarly following training, while locations with intermediate similarity become increasingly decorrelated, both distance-dependent effects that scaled with the compressibility of the input features. Using virtual reality and 2-photon functional calcium imaging in head-fixed mice, we recorded the simultaneous activity of thousands of hippocampal neurons during virtual exploration to test these predictions. Varying the compressibility of sensory information in the environment produced systematic changes in place cell properties that reflected the changing input statistics, consistent with the theory. We similarly identified representational plasticity during learning, which produced a distance-dependent exchange between compression and pattern separation. These results motivate a more domain-general interpretation of hippocampal computation, one that is naturally compatible with earlier theories on the circuit's importance for episodic memory formation. Work done in collaboration with James Priestley, Lorenzo Posani, Marcus Benna, Attila Losonczy.

When a neuron breaks silence, it can emit action potentials in a number of patterns. Some responses are so sudden and intense that electrophysiologists felt the need to single them out, labeling action potentials emitted at a particularly high frequency with a metonym – bursts. Is there more to bursts than a figure of speech? After all, sudden bouts of high-frequency firing are expected to occur whenever inputs surge. In this talk, I will discuss the implications of seeing the neural code as having three syllables: silences, spikes and bursts. In particular, I will describe recent theoretical and experimental results that implicate bursting in the implementation of top-down attention and the coordination of learning.

Neurons in the primary visual cortex (V1) of rodents are selective to the orientation of the stimulus, as in other mammals such as cats and monkeys. However, in contrast with those species, their neurons display a very different type of spatial organization. Instead of orientation maps they are organized in a “salt and pepper” pattern, where adjacent neurons have completely different preferred orientations. This structure has motivated both experimental and theoretical research with the objective of determining which aspects of the connectivity patterns and intrinsic neuronal responses can explain the observed behavior. These analysis have to take into account also that the neurons of the thalamus that send their outputs to the cortex have more complex responses in rodents than in higher mammals, displaying, for instance, a significant degree of orientation selectivity. In this talk we present work showing that a random feed-forward connectivity pattern, in which the probability of having a connection between a cortical neuron and a thalamic neuron depends only on the relative distance between them is enough explain several aspects of the complex phenomenology found in these systems. Moreover, this approach allows us to evaluate analytically the statistical structure of the thalamic input on the cortex. We find that V1 neurons are orientation selective but the preferred orientation of the stimulus depends on the spatial frequency of the stimulus. We disentangle the effect of the non circular thalamic receptive fields, finding that they control the selectivity of the time-averaged thalamic input, but not the selectivity of the time locked component. We also compare with experiments that use reverse correlation techniques, showing that ON and OFF components of the aggregate thalamic input are spatially segregated in the cortex.

The affinity between statistical physics and machine learning has a long history. I will describe the main lines of this long-lasting friendship in the context of current theoretical challenges and open questions about deep learning. Theoretical physics often proceeds in terms of solvable synthetic models, I will describe the related line of work on solvable models of simple feed-forward neural networks. I will highlight a path forward to capture the subtle interplay between the structure of the data, the architecture of the network, and the optimization algorithms commonly used for learning.

The medial entorhinal cortex (MEC) hosts many of the brain’s circuit elements for spatial navigation and episodic memory, operations that require neural activity to be organized across long durations of experience. While location is known to be encoded by a plethora of spatially tuned cell types in this brain region, little is known about how the activity of entorhinal cells is tied together over time. Among the brain’s most powerful mechanisms for neural coordination are network oscillations, which dynamically synchronize neural activity across circuit elements. In MEC, theta and gamma oscillations provide temporal structure to the neural population activity at subsecond time scales. It remains an open question, however, whether similarly coordination occurs in MEC at behavioural time scales, in the second-to-minute regime. In this talk I will show that MEC activity can be organized into a minute-scale oscillation that entrains nearly the entire cell population, with periods ranging from 10 to 100 seconds. Throughout this ultraslow oscillation, neural activity progresses in periodic and stereotyped sequences. The oscillation sometimes advances uninterruptedly for tens of minutes, transcending epochs of locomotion and immobility. Similar oscillatory sequences were not observed in neighboring parasubiculum or in visual cortex. The ultraslow periodic sequences in MEC may have the potential to couple its neurons and circuits across extended time scales and to serve as a scaffold for processes that unfold at behavioural time scales.

A signature feature of cortical circuits is the irregularity of neuronal firing, which manifests itself in the high temporal variability of spiking and the broad distribution of rates. Theoretical works have shown that this feature emerges dynamically in network models if coupling between cells is strong, i.e. if the mean number of synapses per neuron K is large and synaptic efficacy is of order 1/\sqrt{K}. However, the degree to which these models capture the mechanisms underlying neuronal firing in cortical circuits is not fully understood. Results have been derived using neuron models with current-based synapses, i.e. neglecting the dependence of synaptic current on the membrane potential, and an understanding of how irregular firing emerges in models with conductance-based synapses is still lacking. Moreover, at odds with the nonlinear responses to multiple stimuli observed in cortex, network models with strongly coupled cells respond linearly to inputs. In this talk, I will discuss the emergence of irregular firing and nonlinear response in networks of leaky integrate-and-fire neurons. First, I will show that, when synapses are conductance-based, irregular firing emerges if synaptic efficacy is of order 1/\log(K) and, unlike in current-based models, persists even under the large heterogeneity of connections which has been reported experimentally. I will then describe an analysis of neural responses as a function of coupling strength and show that, while a linear input-output relation is ubiquitous at strong coupling, nonlinear responses are prominent at moderate coupling. I will conclude by discussing experimental evidence of moderate coupling and loose balance in the mouse cortex.

Synaptic plasticity is known to be a key player in the brain’s life-long learning abilities. However, due to experimental limitations, the nature of the local changes at individual synapses and their link with emerging network-level computations remain unclear. I will present a numerical, meta-learning approach to deduce plasticity rules from either neuronal activity data and/or prior knowledge about the network's computation. I will first show how to recover known rules, given a human-designed loss function in rate networks, or directly from data, using an adversarial approach. Then I will present how to scale-up this approach to recurrent spiking networks using simulation-based inference.

Understanding Human ability to learn novel concepts from just a few sensory experiences is a fundamental problem in cognitive neuroscience. I will describe a recent work with Ben Sorcher and Surya Ganguli (PNAS, October 2022) in which we propose a simple, biologically plausible, and mathematically tractable neural mechanism for few-shot learning of naturalistic concepts. We posit that the concepts that can be learned from few examples are defined by tightly circumscribed manifolds in the neural firing-rate space of higher-order sensory areas. Discrimination between novel concepts is performed by downstream neurons implementing ‘prototype’ decision rule, in which a test example is classified according to the nearest prototype constructed from the few training examples. We show that prototype few-shot learning achieves high few-shot learning accuracy on natural visual concepts using both macaque inferotemporal cortex representations and deep neural network (DNN) models of these representations. We develop a mathematical theory that links few-shot learning to the geometric properties of the neural concept manifolds and demonstrate its agreement with our numerical simulations across different DNNs as well as different layers. Intriguingly, we observe striking mismatches between the geometry of manifolds in intermediate stages of the primate visual pathway and in trained DNNs. Finally, we show that linguistic descriptors of visual concepts can be used to discriminate images belonging to novel concepts, without any prior visual experience of these concepts (a task known as ‘zero-shot’ learning), indicated a remarkable alignment of manifold representations of concepts in visual and language modalities. I will discuss ongoing effort to extend this work to other high level cognitive tasks.

Neural activity in many sensory systems is organized on low-dimensional manifolds by means of convex receptive fields. Neural codes in these areas are constrained by this organization, as not every neural code is compatible with convex receptive fields. The same codes are also constrained by the structure of the underlying neural network. In my talk I will attempt to provide answers to the following natural questions: (i) How do recurrent circuits generate codes that are compatible with the convexity of receptive fields? (ii) How can we utilize the constraints imposed by the convex receptive field to understand the underlying stimulus space. To answer question (i), we describe the combinatorics of the steady states and fixed points of recurrent networks that satisfy the Dale’s law. It turns out the combinatorics of the fixed points are completely determined by two distinct conditions: (a) the connectivity graph of the network and (b) a spectral condition on the synaptic matrix. We give a characterization of exactly which features of connectivity determine the combinatorics of the fixed points. We also find that a generic recurrent network that satisfies Dale's law outputs convex combinatorial codes. To address question (ii), I will describe methods based on ideas from topology and geometry that take advantage of the convex receptive field properties to infer the dimension of (non-linear) neural representations. I will illustrate the first method by inferring basic features of the neural representations in the mouse olfactory bulb.

Behavioral studies have demonstrated that certain task features reliably enhance classification performance for challenging visual stimuli. These include extended image presentation time and the valid cueing of attention. Here, I will show how convolutional neural networks can be used as a model of the visual system that connects neural activity changes with such performance changes. Specifically, I will discuss how different anatomical forms of recurrence can account for better classification of noisy and degraded images with extended processing time. I will then show how experimentally-observed neural activity changes associated with feature attention lead to observed performance changes on detection tasks. I will also discuss the implications these results have for how we identify the neural mechanisms and architectures important for behavior.