Aug 2026· Journal of Computational Neuroscience· 0 citations· 64 references
Medicine
TL;DR
Training RNNs on delayed decision-making tasks with progressively increasing temporal demands shows that temporal and decision-related computations can emerge through multiple dynamical regimes, while maintaining structured low-dimensional representations and comparable behavioural performance, mirroring biological principles of degeneracy and functional redundancy.
Abstract
How do recurrent neural networks (RNNs) internally represent elapsed time to initiate responses after learned delays? To address this question, we trained RNNs on delayed decision-making tasks with progressively increasing temporal demands, including binary decisions, context-dependent decisions, and perceptual integration. We analyzed trained networks using connectivity statistics, eigenvalue spectra, readout alignment, and low-dimensional population trajectories. Across tasks, networks converged to qualitatively distinct but behaviourally comparable dynamical solutions, including oscillatory and non-oscillatory (ramping/decaying) regimes, consistent with solution degeneracy. Population activity was well approximated by a low-dimensional subspace and distributed across recurrent units rather than localized to individual neurons. Readout alignment was strongly epoch-dependent: as required by the near-zero target output during that epoch, activity evolved largely in the readout-null subspace prior to response generation, and became increasingly aligned with the output dimension near decision time. In sign-symmetric tasks, trained networks preserved exact sign-flip equivariance inherited from architecture and training symmetry. Together, these results show that temporal and decision-related computations can emerge through multiple dynamical regimes, while maintaining structured low-dimensional representations and comparable behavioural performance, mirroring biological principles of degeneracy and functional redundancy.
Decision formation is commonly described as the accumulation of noisy evidence in a low-dimensional decision variable, but it remains unclear how this latent computation is implemented by heterogeneous neural responses. Here, we propose that ramping and sequentially firing neurons form complementary population codes for the same decision variable. Inspired by Laplace-domain neural representations of time, exponential receptive fields in a ramping population give rise to a translatable edge-like activity profile; localized receptive fields in a sequential population give rise to an aligned bump-like profile. We construct a continuous attractor neural network that dynamically maintains these complementary representations while implementing evidence accumulation along a shared latent manifold. At the behavioral level, simulations show that the circuit closely reproduces the single-trial trajectories, choice probabilities, and reaction-time statistics of a standard diffusion decision model while generating heterogeneous ramping and sequential neural responses. Our framework connects latent behavioral dynamics, population geometry, and recurrent circuit mechanisms. More broadly, it provides a circuit-level realization of computation in the Laplace domain that may support the representation and updating of continuous cognitive variables across decision making, timing, memory, and spatial cognition.
Chenyu Wang, Rui Cao, Marc W. Howard· bioRxiv· 0 citations
Neural encoding, decoding, and representation-learning approaches have substantially advanced our ability to predict sensory, behavioral, and cognitive variables from neural population activity. At the same time, dynamical systems approaches increasingly model neural computation as the evolution of latent population states governed by recurrent interactions and structured state transitions. Although these frameworks are often presented as competing paradigms, both can successfully reproduce neural observations while still failing to uniquely identify the computational mechanisms implemented by biological circuits. This perspective argues that this reflects a central unresolved challenge in systems neuroscience: observational recordings alone often provide insufficient constraints for distinguishing mechanistically valid neural dynamics from observationally equivalent alternatives. Accordingly, this perspective proposes a unifying framework integrating representational models, latent neural dynamics, identifiability analysis, and perturbation-based validation within a common mechanistic perspective. First, neural representations are discussed as potentially emerging from temporally localized projections of underlying latent dynamical processes evolving on low-dimensional manifolds. Second, recent advances in learning neural evolution operators using recurrent neural networks, latent state-space models, and dynamical system reconstruction methods are reviewed. Third, it is argued that latent trajectories and predictive performance alone do not guarantee mechanistic validity because multiple latent organizations and evolution operators may remain observationally equivalent despite implying distinct causal mechanisms. Finally, perturbation, intervention, and closed-loop neural interfaces are discussed as additional causal constraints capable of falsifying candidate dynamical explanations under targeted manipulation. Across these four principles, the central challenge in modern neuroscience is framed not simply as decoding neural activity or reconstructing latent trajectories, but as determining which inferred dynamical operators remain predictive under intervention and how perturbation can reduce the admissible class of observationally equivalent candidate mechanisms. From this perspective, evolution operators become experimentally testable hypotheses rather than purely descriptive latent models. Integrating latent dynamical modeling with perturbation-based validation may therefore support a transition from prediction-oriented neuroscience toward perturbation-constrained mechanistic dynamical neuroscience.
Armin Hakkak Moghadam Torbati· Journal of Neural Engineerin...· 0 citations
Dimensionality reduction has proven powerful for identifying neural manifolds, which are low-dimensional structures underlying high-dimensional neural activity. These low-dimensional representations have improved the interpretability of population-level coding. Yet whether such low-dimensional representations are biologically relevant and confer functional advantages in learning systems, or merely reflect neuron-level activity, remains contested in neuroscience. We show that an explicit information bottleneck forcing a recurrent neural network to learn a low-dimensional representation is necessary for rotational and out-of-distribution generalisation in a time-series prediction task. Using information-theoretic measures of causal emergence, we characterise the dynamics of this representation across the memorisation-to-generalisation transition, finding a non-monotonic trajectory which shows an initial decrease, a minimum, and a subsequent rise to a maximum, even as prediction loss falls monotonically. This trajectory scales with task complexity, and the magnitude of emergent structure reliably predicts generalisation performance. Analysis of CA1 hippocampal activity in mice learning an alternating maze task reveals analogous non-monotonic emergence dynamics that track behavioural performance. Together, these findings indicate that the ability of neural networks to learn compact, distributed and emergent representations confers a functional advantage for generalisation, supporting a causal role for learned representations in cognition.
The ability to extract and exploit temporal structure across diverse tasks is central to human cognition. Neuroscientists have typically relied on recurrent neural networks (RNNs) trained with backpropagation through time (BPTT) when modelling neural and behavioural processes such as decision-making and motor control. However, this algorithm has limited biological plausibility, hence the computational principles underlying efficient learning of temporal dependencies remain unresolved. Here, we investigate temporal predictive coding (tPC), a recently proposed framework that extends predictive coding to the temporal domain while preserving local Hebbian update rules. We analyse and extend tPC to establish its relationship with several influential computational models of learning in RNNs, including BPTT, reservoir computing, and eligibility propagation (e-prop). We first demonstrate a functional equivalence between tPC and tBPTT1, a variant of BPTT in which gradients are propagated only one time step into the past. We then show that tPC can leverage reservoir dynamics to encode short-range temporal context, and simultaneously sculpt neural trajectories in state space to support downstream readout. We further demonstrate that hierarchical recurrent dynamics can facilitate learning of more complex temporal dependencies, while additionally conferring robustness to strong distractors. Finally, we show that tPC networks can be augmented with biologically inspired eligibility traces to solve temporally extended context-dependent tasks. Together, these results reveal that relatively simple recurrent networks governed by local plasticity can support temporal learning in more complex settings than previously appreciated. Author summary To navigate the world, our brains must constantly track how events unfold over time, whether we are predicting the next word in a sentence or timing a tennis swing. Neuroscientists often use artificial neural networks to study how the brain learns these sequences. However, these models are often not biologically realistic, since they adjust neuronal connections using information that individual neurons would not have access to in the brain. Here, we examine an alternative framework called temporal predictive coding, which is designed to better reflect how brain networks may be organised. We show that these biologically inspired networks are surprisingly powerful: they can remember recent context and, when arranged in layers, can learn increasingly complex patterns. We also show that they work better when equipped with a memory-like mechanism that helps neurons link recent events across time. Overall, our findings suggest that relatively simple, biologically inspired networks may capture how we process sequential events better than previously thought.
Julian Ng-Kee-Kwong, Mufeng Tang, T. Akam et al.· bioRxiv· 0 citations
SpikeWorld, a 1.45M-parameter sparse spiking model jointly trained for heterogeneous sensory prediction, semantics, image-text binding and action-conditioned dynamics, is introduced, showing that the contribution is not superior linear identification, but its integration with a frozen multimodal spiking checkpoint.
The activity of large neuronal populations is often confined to low-dimensional manifolds that can drift over time, posing a challenge for learning rules that assume stable, full-rank representations. Here, we introduce SPLiT (Synaptic Projection Learning with intrinsic Tracking), a synaptic plasticity rule for recurrent neural networks that combines unsupervised manifold tracking with supervised learning. SPLiT uses an online Oja rule to continuously estimate the intrinsic low-dimensional activity subspace and performs normalized least-mean-squares learning in manifold coordinates. In this way, SPLiT keeps synaptic weights aligned with evolving population dynamics. Using a rate-based recurrent network, we show that SPLiT reliably learns time-varying target signals under both constrained dynamics, where activity is restricted to a fixed low-dimensional manifold, and unconstrained dynamics exhibiting changes in the dominant activity subspace. We show that manifold-constrained learning with SPLiT yields faster convergence, less sensitivity to recurrent gain, smaller weight updates, and is more robust to noisy teaching signals. Analytical results show that SPLiT learns the optimal decoder projected onto the instantaneous principal subspace and maintains bounded error under drift. Together, these findings provide a mechanistic account of how synaptic plasticity can leverage the structure of neural manifolds to enable stable and efficient supervised learning despite representational drift.