Speaker
Description
Sensorimotor systems in humans and other animals sustain complex and adaptive behavior by continuously integrating noisy sensory feedback with internal states and behavioral goals. This behavior is remarkably reliable despite variability affecting every stage of the perception-action loop, from stochastic body and environmental dynamics to sensory and motor noise, as well as fluctuations in internal representations. Stochastic optimal control has long provided a normative framework for understanding how biological agents should act under uncertainty, but it leaves open whether and how these computational principles can be implemented in neural circuits.
Recent theoretical work on stochastic optimal control has emphasized the role of internal variability in shaping control strategies, suggesting that robust behavior may require internal dynamics that protect motor outputs from fluctuations in the controller. Here, we ask whether this principle naturally emerges in recurrent neural networks trained on sensorimotor control tasks. We find that increasing internal noise drives a systematic reorganization of neural dynamics: neural activity becomes increasingly oblique to the motor readout, while trial-to-trial variability is redistributed away from output-relevant directions. This echoes recent work on RNNs showing that noise injected into recurrent dynamics can promote robust oblique dynamics. Here, however, internal noise is not introduced as a regularizer for learning, but as a biologically motivated component of the sensorimotor loop, capturing fluctuations in the internal representations that transform sensory feedback into action.
Finally, we show that the trained networks admit a low-rank dynamical description that captures the main control-relevant modes and explains how robust behavior can coexist with large internal fluctuations. Together, these results show that computational principles of robust motor control under variability, predicted by stochastic optimal-control theory, can naturally emerge in trained recurrent neural circuits. More broadly, they suggest that oblique recurrent dynamics may provide a neural implementation of robust control under internal variability.
| Preferred Presentation | Poster Presentation |
|---|