A New Look at Low-Rank Recurrent Neural Networks

ICLR 2025 Conference Submission12310 Authors

27 Sept 2024 (modified: 21 Nov 2024)ICLR 2025 Conference SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: low-rank rnn, computational neuroscience, dynamical systems, neural dynamics
TL;DR: We introduce an approach to train low-rank RNNs to implement a desired dynamical system, clarify limits on expressivity of such models, and offer insights into how input-driven RNNs produce trajectories observed in neuroscience tasks.
Abstract: Low-rank recurrent neural networks (RNNs) have recently gained prominence as a framework for understanding how neural systems solve complex cognitive tasks. However, training and interpreting these networks remains an important open problem. Here we address these challenges by adopting a view of low-rank RNNs as parametrizing a low-dimensional ordinary differential equation (ODE) using a set of nonlinear basis functions. This perspective, which arises from an approach known as the ``neural engineering framework'', reveals that low-rank RNNs are equivalent to neural ODEs with a single hidden layer. We show that training a low-rank RNN to implement a particular dynamical system can thus be formalized as least-squares regression in a random basis. This allows us to propose a new method for finding the smallest RNN capable of implementing a dynamical system using a variant of orthogonal matching pursuit. More generally, our perspective clarifies limits on the expressivity of low-rank RNNs, such as the fact that without inputs, a low-rank RNN with sigmoidal nonlinearity can only implement odd-symmetric functions. We delve further into the role of inputs in shaping network dynamics and show that RNNs can produce identical trajectories using a wide variety of static or time-varying dynamics; this highlights the importance of perturbations for inferring dynamics from observed neural trajectories. Finally, we highlight the usefulness of our framework by comparing to RNNs trained using backprop-through-time on neuroscience-inspired tasks, showcasing that our method achieves faster and more accurate learning with smaller networks than gradient-based training.
Supplementary Material: zip
Primary Area: applications to neuroscience & cognitive science
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12310
Loading