Many natural systems, including neural circuits involved in decision making, are modeled as high-dimensional dynamical systems with multiple stable states. While existing analytical tools primarily describe behavior near stable equilibria, characterizing separatrices -- the manifolds that delineate boundaries between different basins of attraction -- remains challenging, particularly in high-dimensional settings. Here, we introduce a numerical framework leveraging Koopman Theory combined with Deep Neural Networks to effectively characterize separatrices. Specifically, we approximate Koopman Eigenfunctions (KEFs) associated with real positive eigenvalues, which vanish precisely at the separatrices. Utilizing these scalar KEFs, optimization methods efficiently locate separatrices even in complex systems. We demonstrate our approach on synthetic benchmarks, ecological network models, and high-dimensional recurrent neural networks trained on either neuroscience-inspired tasks or fit to real neural data. Moreover, we illustrate the practical utility of our method by designing optimal perturbations that can shift systems across separatrices, enabling predictions relevant to optogenetic stimulation experiments in neuroscience. Our code is available on [GitHub](https://github.com/KabirDabholkar/separatrixLocator), and we share an interactive description of the work and its extensions in a [UniReps blog](https://unireps.github.io/blog/2025/Separatrix-Locator/).

Recurrent neural networks (RNNs) trained on neuroscience-inspired tasks offer powerful models of brain computation. However, typical training paradigms rely on open-loop, supervised settings, whereas real-world learning unfolds in closed-loop environments. Here, we develop a mathematical theory describing the learning dynamics of linear RNNs trained in closed-loop contexts. We first demonstrate that two otherwise identical RNNs, trained in either closed- or open-loop modes, follow markedly different learning trajectories. To probe this divergence, we analytically characterize the closed-loop case, revealing distinct stages aligned with the evolution of the training loss. Specifically, we show that the learning dynamics of closed-loop RNNs, in contrast to open-loop ones, are governed by an interplay between two competing objectives: short-term policy improvement and long-term stability of the agent-environment interaction. Finally, we apply our framework to a realistic motor control task, highlighting its broader applicability. Taken together, our results underscore the importance of modeling closed-loop dynamics in a biologically plausible setting.

Attractors play a key role in a wide range of processes including learning and memory. Due to recent innovations in recording methods, there is increasing evidence for the existence of attractor dynamics in the brain. Yet, our understanding of how these attractors emerge or disappear in a biological system is lacking. By following the spontaneous network bursts of cultured cortical networks, we are able to define a vocabulary of spatiotemporal patterns and show that they function as discrete attractors in the network dynamics. We show that electrically stimulating specific attractors eliminates them from the spontaneous vocabulary, while they are still robustly evoked by the electrical stimulation. This seemingly paradoxical finding can be explained by a Hebbian-like strengthening of specific pathways into the attractors, at the expense of weakening non-evoked pathways into the same attractors. We verify this hypothesis and provide a mechanistic explanation for the underlying changes supporting this effect.

Neural oscillations are ubiquitously observed in many brain areas. One proposed functional role of these oscillations is that they serve as an internal clock, or 'frame of reference'. Information can be encoded by the timing of neural activity relative to the phase of such oscillations. In line with this hypothesis, there have been multiple empirical observations of such phase codes in the brain. Here we ask: What kind of neural dynamics support phase coding of information with neural oscillations? We tackled this question by analyzing recurrent neural networks (RNNs) that were trained on a working memory task. The networks were given access to an external reference oscillation and tasked to produce an oscillation, such that the phase difference between the reference and output oscillation maintains the identity of transient stimuli. We found that networks converged to stable oscillatory dynamics. Reverse engineering these networks revealed that each phase-coded memory corresponds to a separate limit cycle attractor. We characterized how the stability of the attractor dynamics depends on both reference oscillation amplitude and frequency, properties that can be experimentally observed. To understand the connectivity structures that underlie these dynamics, we showed that trained networks can be described as two phase-coupled oscillators. Using this insight, we condensed our trained networks to a reduced model consisting of two functional modules: One that generates an oscillation and one that implements a coupling function between the internal oscillation and external reference. In summary, by reverse engineering the dynamics and connectivity of trained RNNs, we propose a mechanism by which neural networks can harness reference oscillations for working memory. Specifically, we propose that a phase-coding network generates autonomous oscillations which it couples to an external reference oscillation in a multi-stable fashion.

How does a single interconnected neural population perform multiple tasks, each with its own dynamical requirements? The relation between task requirements and neural dynamics in Recurrent Neural Networks (RNNs) has been investigated for single tasks. The forces shaping joint dynamics of multiple tasks, however, are largely unexplored. In this work, we first construct a systematic framework to study multiple tasks in RNNs, minimizing interference from input and output correlations with the hidden representation. This allows us to reveal how RNNs tend to share attractors and reuse dynamics, a tendency we define as the "simplicity bias". We find that RNNs develop attractors sequentially during training, preferentially reusing existing dynamics and opting for simple solutions when possible. This sequenced emergence and preferential reuse encapsulate the simplicity bias. Through concrete examples, we demonstrate that new attractors primarily emerge due to task demands or architectural constraints, illustrating a balance between simplicity bias and external factors. We examine the geometry of joint representations within a single attractor, by constructing a family of tasks from a set of functions. We show that the steepness of the associated functions controls their alignment within the attractor. This arrangement again highlights the simplicity bias, as points with similar input spacings undergo comparable transformations to reach the shared attractor. Our findings propose compelling applications. The geometry of shared attractors might allow us to infer the nature of unknown tasks. Furthermore, the simplicity bias implies that without specific incentives, modularity in RNNs may not spontaneously emerge, providing insights into the conditions required for network specialization.

In recent years Recurrent Neural Networks (RNNs) were successfully used to model the way neural activity drives task-related behavior in animals, operating under the implicit assumption that the obtained solutions are universal. Observations in both neuroscience and machine learning challenge this assumption. Animals can approach a given task with a variety of strategies, and training machine learning algorithms introduces the phenomenon of underspecification. These observations imply that every task is associated with a space of solutions. To date, the structure of this space is not understood, limiting the approach of comparing RNNs with neural data.Here, we characterize the space of solutions associated with various tasks. We first study a simple two-neuron network on a task that leads to multiple solutions. We trace the nature of the final solution back to the network’s initial connectivity and identify discrete dynamical regimes that underlie this diversity. We then examine three neuroscience-inspired tasks: Delayed discrimination, Interval discrimination, and Time reproduction. For each task, we find a rich set of solutions. One layer of variability can be found directly in the neural activity of the networks. An additional layer is uncovered by testing the trained networks' ability to extrapolate, as a perturbation to a system often reveals hidden structure. Furthermore, we relate extrapolation patterns to specific dynamical objects and effective algorithms found by the networks. We introduce a tool to derive the reduced dynamics of networks by generating a compact directed graph describing the essence of the dynamics with regards to behavioral inputs and outputs. Using this representation, we can partition the solutions to each task into a handful of types and show that neural features can partially predict them.Taken together, our results shed light on the concept of the space of solutions and its uses both in Machine learning and in Neuroscience.

Empirical estimates of the dimensionality of neural population activity are often much lower than the population size. Similar phenomena are also observed in trained and designed neural network models. These experimental and computational results suggest that mapping low-dimensional dynamics to high-dimensional neural space is a common feature of cortical computation. Despite the ubiquity of this observation, the constraints arising from such mapping are poorly understood. Here we consider a specific example of mapping low-dimensional dynamics to high-dimensional neural activity-the neural engineering framework. We analytically solve the framework for the classic ring model-a neural network encoding a static or dynamic angular variable. Our results provide a complete characterization of the success and failure modes for this model. Based on similarities between this and other frameworks, we speculate that these results could apply to more general scenarios.

Training recurrent neural networks (RNNs) on low-dimensional tasks has been widely used to model functional biological networks. However, the solutions found by learning and the effect of initial connectivity are not well understood. Here, we examine RNNs trained using gradient descent on different tasks inspired by the neuroscience literature. We find that the changes in recurrent connectivity can be described by low-rank matrices. This observation holds even in the presence of random initial connectivity, although this initial connectivity has full rank and significantly accelerates training. To understand the origin of these observations, we turn to an analytically tractable setting: training a linear RNN on a simpler task. We show how the low-dimensional task structure leads to low-rank changes to connectivity, and how random initial connectivity facilitates learning. Altogether, our study opens a new perspective to understand learning in RNNs in light of low-rank connectivity changes and the synergistic role of random initialization.

Biological networks are often heterogeneous in their connectivity pattern, with degree distributions featuring a heavy tail of highly connected hubs. The implications of this heterogeneity on dynamical properties are a topic of much interest. Here we show that interpreting topology as a feedback circuit can provide novel insights on dynamics. Based on the observation that in finite networks a small number of hubs have a disproportionate effect on the entire system, we construct an approximation by lumping these nodes into a single effective hub, which acts as a feedback loop with the rest of the nodes. We use this approximation to study dynamics of networks with scale-free degree distributions, focusing on their probability of convergence to fixed points. We find that the approximation preserves convergence statistics over a wide range of settings. Our mapping provides a parametrization of scale free topology which is predictive at the ensemble level and also retains properties of individual realizations. Specifically, outgoing hubs have an organizing role that can drive the network to convergence, in analogy to suppression of chaos by an external drive. In contrast, incoming hubs have no such property, resulting in a marked difference between the behavior of networks with outgoing vs. incoming scale free degree distribution. Combining feedback analysis with mean field theory predicts a transition between convergent and divergent dynamics which is corroborated by numerical simulations. Furthermore, they highlight the effect of a handful of outlying hubs, rather than of the connectivity distribution law as a whole, on network dynamics.

To be effective in sequential data processing, Recurrent Neural Networks (RNNs) are required to keep track of past events by creating memories. While the relation between memories and the network’s hidden state dynamics was established over the last decade, previous works in this direction were of a predominantly descriptive nature focusing mainly on locating the dynamical objects of interest. In particular, it remained unclear how dynamical observables affect the performance, how they form and whether they can be manipulated. Here, we utilize different training protocols, datasets and architectures to obtain a range of networks solving a delayed classification task with similar performance, alongside substantial differences in their ability to extrapolate for longer delays. We analyze the dynamics of the network’s hidden state, and uncover the reasons for this difference. Each memory is found to be associated with a nearly steady state of the dynamics which we refer to as a ’slow point’. Slow point speeds predict extrapolation performance across all datasets, protocols and architectures tested. Furthermore, by tracking the formation of the slow points we are able to understand the origin of differences between training protocols. Finally, we propose a novel regularization technique that is based on the relation between hidden state speeds and memory longevity. Our technique manipulates these speeds, thereby leading to a dramatic improvement in memory robustness over time, and could pave the way for a new class of regularization methods.

Artificial neural networks, trained to perform cognitive tasks, have recently been used as models for neural recordings from animals performing these tasks. While some progress has been made in performing such comparisons, the evolution of network dynamics throughout learning remains unexplored. This is paralleled by an experimental focus on recording from trained animals, with few studies following neural activity throughout training. In this work, we address this gap in the realm of artificial networks by analyzing networks that are trained to perform memory and pattern generation tasks. The functional aspect of these tasks corresponds to dynamical objects in the fully trained network-a line attractor or a set of limit cycles for the two respective tasks. We use these dynamical objects as anchors to study the effect of learning on their emergence. We find that the sequential nature of learning-one trial at a time-has major consequences for the learning trajectory and its final outcome. Specifically, we show that least mean squares (LMS), a simple gradient descent suggested as a biologically plausible version of the FORCE algorithm, is constantly obstructed by forgetting, which is manifested as the destruction of dynamical objects from previous trials. The degree of interference is determined by the correlation between different trials. We show which specific ingredients of FORCE avoid this phenomenon. Overall, this difference results in convergence that is orders of magnitude slower for LMS. Learning implies accumulating information across multiple trials to form the overall concept of the task. Our results show that interference between trials can greatly affect learning in a learning-rule-dependent manner. These insights can help design experimental protocols that minimize such interference, and possibly infer underlying learning rules by observing behavior and neural activity throughout learning.

Our acoustical environment abounds with repetitive sounds, some of which are related to pitch perception. It is still unknown how the auditory system, in processing these sounds, relates a physical stimulus and its percept. Since, in mammals, all auditory stimuli are conveyed into the nervous system through the auditory nerve (AN) fibers, a model should explain the perception of pitch as a function of this particular input. However, pitch perception is invariant to certain features of the physical stimulus. For example, a missing fundamental stimulus with resolved or unresolved harmonics, or a low and high-level amplitude stimulus with the same spectral content-these all give rise to the same percept of pitch. In contrast, the AN representations for these different stimuli are not invariant to these effects. In fact, due to saturation and non-linearity of both cochlear and inner hair cells responses, these differences are enhanced by the AN fibers. Thus there is a difficulty in explaining how pitch percept arises from the activity of the AN fibers. We introduce a novel approach for extracting pitch cues from the AN population activity for a given arbitrary stimulus. The method is based on a technique known as sparse coding (SC). It is the representation of pitch cues by a few spatiotemporal atoms (templates) from among a large set of possible ones (a dictionary). The amount of activity of each atom is represented by a non-zero coefficient, analogous to an active neuron. Such a technique has been successfully applied to other modalities, particularly vision. The model is composed of a cochlear model, an SC processing unit, and a harmonic sieve. We show that the model copes with different pitch phenomena: extracting resolved and non-resolved harmonics, missing fundamental pitches, stimuli with both high and low amplitudes, iterated rippled noises, and recorded musical instruments.

Recurrent neural networks (RNNs) are useful tools for learning nonlinear relationships between time-varying inputs and outputs with complex temporal dependencies. Recently developed algorithms have been successful at training RNNs to perform a wide variety of tasks, but the resulting networks have been treated as black boxes: their mechanism of operation remains unknown. Here we explore the hypothesis that fixed points, both stable and unstable, and the linearized dynamics around them, can reveal crucial aspects of how RNNs implement their computations. Further, we explore the utility of linearization in areas of phase space that are not true fixed points but merely points of very slow movement. We present a simple optimization technique that is applied to trained RNNs to find the fixed and slow points of their dynamics. Linearization around these slow regions can be used to explore, or reverse-engineer, the behavior of the RNN. We describe the technique, illustrate it using simple examples, and finally showcase it on three high-dimensional RNN examples: a 3-bit flip-flop device, an input-dependent sine wave generator, and a two-point moving average. In all cases, the mechanisms of trained networks could be inferred from the sets of fixed and slow points and the linearized dynamics around them.

The perceptron is a simple supervised algorithm to train a linear classifier that has been analyzed and used extensively. The classifier separates the data into two groups using a decision hyperplane, with the margin between the data and the hyperplane determining the classifier's ability to generalize and its robustness to input noise. Exact results for the maximal size of the separating margin are known for specific input distributions, and bounds exist for arbitrary distributions, but both rely on lengthy statistical mechanics calculations carried out in the limit of infinite input size. Here we present a short analysis of perceptron classification using singular value decomposition. We provide a simple derivation of a lower bound on the margin and an explicit formula for the perceptron weights that converges to the optimal result for large separating margins.

Persistent activity states (attractors), observed in several neocortical areas after the removal of a sensory stimulus, are believed to be the neuronal basis of working memory. One of the possible mechanisms that can underlie persistent activity is recurrent excitation mediated by intracortical synaptic connections. A recent experimental study revealed that connections between pyramidal cells in prefrontal cortex exhibit various degrees of synaptic depression and facilitation. Here we analyze the effect of synaptic dynamics on the emergence and persistence of attractor states in interconnected neural networks. We show that different combinations of synaptic depression and facilitation result in qualitatively different network dynamics with respect to the emergence of the attractor states. This analysis raises the possibility that the framework of attractor neural networks can be extended to represent time-dependent stimuli.

In PLoS Computational Biology, volume 3, issue 2: doi: 10.1371/journal.pcbi.0030035 Equation 28 had incorrect symbols in lines 3, 6, and 7. Here is the correct Equation 28:

Recognizing specific spatiotemporal patterns of activity, which take place at timescales much larger than the synaptic transmission and membrane time constants, is a demand from the nervous system exemplified, for instance, by auditory processing. We consider the total synaptic input that a single readout neuron receives on presentation of spatiotemporal spiking input patterns. Relying on the monotonic relation between the mean and the variance of a neuron's input current and its spiking output, we derive learning rules that increase the variance of the input current evoked by learned patterns relative to that obtained from random background patterns. We demonstrate that the model can successfully recognize a large number of patterns and exhibits a slow deterioration in performance with increasing number of learned patterns. In addition, robustness to time warping of the input patterns is revealed to be an emergent property of the model. Using a leaky integrate-and-fire realization of the readout neuron, we demonstrate that the above results also apply when considering spiking output.