Recurrent neural networks

Riccardo Zecchina presents a learning rule for recurrent neural networks in which nonlinear dendrites create and stabilize task-relevant attractors within a representation manifold. Learning is local in both space and time. On the same network architecture, the scheme matches the performance of backpropagation through time while taking a fundamentally different path through weight space. Statistical-physics analysis explains how dendritic gating and the geometry of the representation manifold support this alternative mechanism for learning.

…representations emerge in recurrent neural networks trained to predict sequences of sensory observations. While spatially tuned units reliably arise, they are not sufficient to form a cognitive map. Instead, map-like representations emerge when recurrent dynamics support multi-step prediction, yielding a population-level encoding of environmental geometry. On

…storing continuous-valued variables in recurrent neural states for indefinitely long time intervals. Unfortunately, continuous attractors suffer from severe structural instability in general---they are destroyed by most infinitesimal changes of the dynamical law that defines them. This fragility may limit their utility especially in biological systems as their recurrent dynamics are subject to constant perturbations. We indeed observe that the bifurcations from continuous attractors in theoretical neuroscience models

During development, neural circuits are shaped continuously as we learn to control our bodies, with the goal of producing neural dynamics that enable the rich repertoire of behaviors we perform with our limbs. However, the nature of the teaching signal underlying this normative learning process remains elusive. Here, we test two well-established and biologically plausible theories---supervised learning (SL) and reinforcement learning (RL)---that could explain how neural circuits develop the capacity for skilled movements. We trained recurrent neural networks to control a biomechanical model of

…computational primitives in artificial neural networks remains a significant challenge, suggesting that biological systems may possess structural priors which enhance this capability. In this study, we examine the mechanisms by which network structure facilitates flexible compositions of computational primitives. To investigate this question, we analyze recurrent neural networks (RNNs) that perform a temporal vector additio

Recurrent neural networks (RNNs) have increasingly been employed to model cortical function, but conventional architectures lack biological realism. Notably, they often ignore two key biological principles: i) Dale’s Law, which constrains neurons to either be strictly excitatory or inhibitory, and ii) Sparse, structured connectivity motifs observed in the brain. Incorporating these constraints however poses challenges, as they often affect model performance negatively. Our work presents a novel approach that integrates Dale’s backpropagation, a sign-constrained training method, with topologica

…is fundamentally challenging for recurrent neural networks (RNNs). While much prior work has explored why RNNs struggle to learn long timescales and how to mitigate this, we still lack a clear understanding of the dynamics involved when RNNs learn long timescales via gradient descent. Here we build a mathematical theory of the learning dynamics of linear RNNs trained to integrate white noise. We show that when the initial recurrent weights are small, the dynamics of lea

…experience shapes learning, we trained Recurrent Neural Networks (RNNs) to model a complex, context-dependent odor timing task, using constraints derived from prior reports detailing mouse behavior and shaping procedures. The task, a temporal variant of the Delay Non-match

…for finding structure in heterogeneous neural activity. When applied to neural response patterns in a population state space, where each axis represents activity of one neuron, dimensionality reduction methods often reveal low-dimensional manifolds encoding behaviorally relevant variables. Alternatively, when the same responses are viewed in a selectivity space, where each axis represents a response feature (e.g., task condition or time), dimensionality reduction methods can

Targeted photostimulation optogenetics offers unprecedented precision in perturbing neural circuits in action. However, due to the complexity of neural responses, perturbing neurons chosen randomly or based on their task selectivity often produces weak or counter-intuitive effects. Recurrent neural networks (RNNs) trained to replicate neural activity and task behavior hold promise for guiding the design of targeted perturbations. Yet, it remains uncertain whether low-dimensional neural responses during cognitive tasks sufficiently constrain high-dimensional models to predict outcomes of patter

Models of recurrent neural networks provide fundamental insights about the interplay between structure and dynamics in biological neural networks. Most previous work focused on networks with random or simple connectivity structures, yet experiments show that overrepresented motifs shape the patterns of network activity. A theory relating such complex connectivity structures to network dynamics and function is lacking. Here, we show that third and higher-order correlations in synaptic connectivities have a large impact on neuronal dynamics. Strong triangular cyclic correlations in a network su

…type of compositional learning with neural systems. Here we leverage recent work on compositional computations in multitasking Recurrent Neural Networks (RNNs) to construct a neural model of compositional learning. Pa

Task-trained recurrent neural networks (RNNs) are versatile models of dynamical processes widely used in machine learning and neuroscience. While RNNs are easily trained to perform a wide range of tasks, the nature and extent of the degeneracy in the resultant solutions (i.e., the variability across trained RNNs) remain poorly understood. Here, we provide a unified framework for analyzing degeneracy across three levels: behavior, neural dynamics, and weight space. We analyzed RNNs trained on diverse tasks across machine learning and neuroscience domains, including N-bit flip-flop, sine wave ge

…understanding of how the dynamics of neural circuits give rise to their observed function. A popular approach towards this end is to train recurrent neural networks (RNNs) to reproduce experimental recordings of neural activity. These trained RNNs are then treated as surrogate models of biological neural circuits, whose properties can be dissected via dynamical systems analysis. How reliable are the mechanistic insights derived from this procedure? While recent advances in population-level recording technologies have allowed sim

…variability to probe the geometry of neural dynamics (NDs) for movements across different postures and extents. To address this confound,

Two striatal subregions, the Dorsolateral Striatum (DLS) and the Dorsomedial Striatum (DMS) are thought to implement habitual and flexible, goal-directed behavior respectively. The DMS has been shown to be involved in tasks such as gating of information into working memory and accumulating evidence before making a decision. How the DMS produces these behaviors is less understood than the DLS because, unlike habitual motor tasks used to study the DLS, tasks used to study the DMS require more management of internal mental states, an inherently unobservable process. In this work, we build a compu

Artifical neural network (ANN) approaches to motor control commonly adopt generic architectures like fully connected multi-layer perceptrons (MLPs) and recurrent neural networks (RNNs). As such architectures contain few inductive biases, it is common to require priors in the form of rewards, training curricula, imitation data, or trajectory generators. In nature, animals are born with priors in the form of their nervous system's architecture, which has been shaped by evolution to confer innate ability and efficient learning. For instance, a horse can walk within hours of birth and quickly impr

…to link connectivity features of large neural networks to their emergent dynamics. Nevertheless, how to integrate connectomics data and neural recordings to determine the function of a neural system remains challenging, given the physiological properties of individual neurons that remain unconstrained. To address this, we developed a theory that describes the space of solutions when inferring activit

Integrating task-relevant information into neural representations is essential for both biological and artificial intelligence. However, current experimental limitations in tracking synaptic weight changes make it challenging to understand learning mechanisms in animals. To address this, we propose a framework based on neural manifold geometry to study representation learning directly from neural activity. A neural manifold represents the collection of population responses to the same task condition (e.g., a cat manifold for responses to images of cats) and can be estimated from recorded activ

…dynamics, and computation in recurrent neural networks (RNNs), in particular those that can be created via biological learning rules. To date, most work on RNN training has focused on either efficient but non-biological rules (e.g., backprop, FORCE) or biologically plausible rules with poor learning efficiency (e.g., Hebbian learning). However, some of the fastest, strongest learning rules thought to act in vivo are not Hebbian, for instance in the fly mushroom body, where plasticity depends on high-dimensional presynaptic activity

Training recurrent neural networks (RNNs) to perform neuroscience tasks can be challenging. Unlike in machine learning where any architectural modification of an RNN (e.g. GRU or LSTM) is acceptable if it facilitates training, the RNN models trained as models of brain dynamics are subject to plausibility constraints that fundamentally exclude the usual machine learning hacks. The “vanilla” RNNs commonly used in computational neuroscience find themselves plagued by ill-conditioned loss surfaces that complicate training and significantly hinder our capacity to investigate the brain dynamics unde

…heavy-tailed distributions influence neural dynamics is crucial for developing more biologically realistic models and theories that better capture the brain's properties. To bridge this gap, we provide both theoretical and computational insights into the dynamics shaped by heavy-tailed synaptic weights through mathematical derivations and case studies in recurrent neural

…is accumulated or maintained by recurrently connected circuits in the brain. In order to guide the study of how structures of knowledge are accumulated and maintained to support optimal foraging decisions, we reverse-engineered recurrent neural networks (RNNs) trained to perform dynamic foraging tasks. We highlight two variants of the task --- a highly structured block-switch variant and a weakly structured random walk variant --- and investigate how network dynamics are shaped by

Associative learning relies on the contingency between stimuli and outcomes, yet the neural mechanisms instantiating contingency remain unclear. We investigated dopamine activity in the ventral striatum of mice performing a Pavlovian contingency degradation task. Both anticipatory licking behavior and dopamine responses to a conditioned stimulus diminished when additional rewards were delivered without cues (uncued) but remained consistent when the same additional rewards were paired with a new cue. These findings challenge traditional contingency-based accounts and the recent causal learning