A biology-inspired recurrent oscillator network for computations in high-dimensional state space

  • Felix Effenberger (Stealth Silicon Valley Startup, San Francisco, USA)
A3 01 (Sophus-Lie room)


Whether oscillations in biological neuronal networks are merely a byproduct of neuronal interactions or serve computational purposes continues to be a topic of active discussion. Here, we report on how the inclusion of hallmark features of the cerebral cortex such as the presence of oscillatory units, heterogeneity, synaptic delays, and modularity into recurrent neural networks (RNNs) simulated in silico influences their performance on common pattern recognition tasks when trained with a gradient-based learning rule. We find that our RNNs composed of damped harmonic oscillators (DHOs) learn to desynchronize their activity to produce high-dimensional representations of stimuli, and by leveraging non-linear dynamical effects such as frequency dependent gain modulation form a computational substrate that vastly outperforms state of the art gated RNN architectures (GRU, LSTM) in learning speed, task performance, and noise resiliency. Analysis of the structure and dynamics of our networks provides a posteriori explanations for a number of physiological phenomena whose function so far has been elusive or has given rise to controversial discussions.

Katharina Matschke

MPI for Mathematics in the Sciences Contact via Mail