Transition to chaos separates learning regimes and relates to measure of consciousness in recurrent neural networks DOI Creative Commons
D. Mastrovito, Yuhan Helena Liu, Łukasz Kuśmierz

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: May 15, 2024

Recurrent neural networks exhibit chaotic dynamics when the variance in their connection strengths exceed a critical value. Recent work indicates also modulates learning strategies; learn "rich" representations initialized with low coupling and "lazier" solutions larger variance. Using Watts-Strogatz of varying sparsity, structure, hidden weight variance, we find that strength dividing from ordered differentiates rich lazy strategies. Training moves both stable closer to edge chaos, richer before transition chaos. In contrast, biologically realistic connectivity structures foster stability over wide range variances. The chaos is reflected measure clinically discriminates levels consciousness, perturbational complexity index (PCIst). Networks high values PCIst learning, suggesting consciousness prior may promote learning. results suggest clear relationship between dynamics, regimes complexity-based measures consciousness.

Language: Английский

Specific connectivity optimizes learning in thalamocortical loops DOI Creative Commons
Kaushik J. Lakshminarasimhan, Marjorie Xie, Jeremy D. Cohen

et al.

Cell Reports, Journal Year: 2024, Volume and Issue: 43(4), P. 114059 - 114059

Published: April 1, 2024

Thalamocortical loops have a central role in cognition and motor control, but precisely how they contribute to these processes is unclear. Recent studies showing evidence of plasticity thalamocortical synapses indicate for the thalamus shaping cortical dynamics through learning. Since signals undergo compression from cortex thalamus, we hypothesized that computational depends critically on structure corticothalamic connectivity. To test this, identified optimal promotes biologically plausible learning synapses. We found projections specialized communicate an efference copy output benefit while communicating modes highest variance working memory tasks. analyzed neural recordings mice performing grasping delayed discrimination tasks communication consistent with predictions. These results suggest orchestrates functionally precise manner structured

Language: Английский

Citations

6

Feedback control of recurrent dynamics constrains learning timescales during motor adaptation DOI
Harsha Gurnani, Weixuan Liu, Bingni W. Brunton

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: May 26, 2024

Abstract Latent dynamical models of the primary motor cortex (M1) have revealed fundamental neural computations underlying control; however, such often overlook impact sensory feedback, which can continually update cortical dynamics and correct for external perturbations. This suggests a critical need to model interaction between feedback intrinsic dynamics. Such would also benefit design brain-computer interfaces (BCIs) that decode activity in real time, where both user learning proficient control require feedback. Here we investigate flexible modulation demonstrate its on BCI task performance short-term learning. By training recurrent network with real-time simple 2D reaching task, analogous cursor control, show how previously reported M1 patterns be reinterpreted as arising from feedback-driven Next, by incorporating adaptive controllers upstream M1, make testable prediction new decoder is facilitated plasticity inputs including remapping beyond connections within M1. input-driven structure determines speed adaptation outcomes, explains continuous form variability. Thus, our work highlights input-dependent latent clarifies constraints arise statistical characteristics activity.

Language: Английский

Citations

1

Transition to chaos separates learning regimes and relates to measure of consciousness in recurrent neural networks DOI Creative Commons
D. Mastrovito, Yuhan Helena Liu, Łukasz Kuśmierz

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: May 15, 2024

Recurrent neural networks exhibit chaotic dynamics when the variance in their connection strengths exceed a critical value. Recent work indicates also modulates learning strategies; learn "rich" representations initialized with low coupling and "lazier" solutions larger variance. Using Watts-Strogatz of varying sparsity, structure, hidden weight variance, we find that strength dividing from ordered differentiates rich lazy strategies. Training moves both stable closer to edge chaos, richer before transition chaos. In contrast, biologically realistic connectivity structures foster stability over wide range variances. The chaos is reflected measure clinically discriminates levels consciousness, perturbational complexity index (PCIst). Networks high values PCIst learning, suggesting consciousness prior may promote learning. results suggest clear relationship between dynamics, regimes complexity-based measures consciousness.

Language: Английский

Citations

0