Abstract
Recurrent neural networks (RNNs) are powerful dynamical models, widely used in machine learning (ML) and neuroscience. Prior theoretical work has focused on RNNs with additive interactions. However, gating, i.e., multiplicative, interactions are ubiquitous in real neurons and also the central feature of the best-performing RNNs in ML. Here, we show that gating offers flexible control of two salient features of the collective dynamics: (i) timescales and (ii) dimensionality. The gate controlling timescales leads to a novel, marginally stable state, where the network functions as a flexible integrator. Unlike previous approaches, gating permits this important function without parameter fine-tuning or special symmetries. Gates also provide a flexible, context-dependent mechanism to reset the memory trace, thus complementing the memory function. The gate modulating the dimensionality can induce a novel, discontinuous chaotic transition, where inputs push a stable system to strong chaotic activity, in contrast to the typically stabilizing effect of inputs. At this transition, unlike additive RNNs, the proliferation of critical points (topological complexity) is decoupled from the appearance of chaotic dynamics (dynamical complexity). The rich dynamics are summarized in phase diagrams, thus providing a map for principled parameter initialization choices to ML practitioners.
6 More- Received 15 March 2021
- Revised 21 June 2021
- Accepted 11 November 2021
DOI:https://doi.org/10.1103/PhysRevX.12.011011
Published by the American Physical Society under the terms of the Creative Commons Attribution 4.0 International license. Further distribution of this work must maintain attribution to the author(s) and the published article’s title, journal citation, and DOI.
Published by the American Physical Society
Physics Subject Headings (PhySH)
Popular Summary
Recurrent neural networks (RNNs) are responsible for impressive advances in modern artificial intelligence, and they are widely used to model collective behavior of neurons in brains. The success of modern RNNs can be largely attributed to one key feature: gating, a multiplicative interaction (also present in real neurons) that controls the flow of information. Despite this spectacular success, it is unclear how gating shapes the behavior of RNNs. We provide a comprehensive theory of gating in RNNs, which gives engineers a road map to push the state of the art in applications and provides neuroscientists with new insights into the collective behavior of neurons.
Gating either suppresses or allows input into a neuron, whether real or artificial. In RNNs, this ability leads to greatly improved information processing abilities. In our work, we study how this affects the dynamics of a model RNN by extending a classic RNN model to include gates that influence timescale and dimensionality. The timescale gate leads to a novel, marginally stable state that can functionally serve as a neural integrator. The dimensionality gate gives rise to an input-induced transition to chaos. The interplay of these dynamical phenomena with inputs allows gated RNNs to serve as flexible memories.
Our work demonstrates the power of physics-based approaches in studying models in machine learning and neuroscience and further invigorates the intense activity at this fertile interface.