Rainer Engelken: Rainer Engelken: Chaos, Stability, and Trainability in Sparse Recurrent Neural Networks: From spiking network chaos to input control and learning dynamics
| When |
Jul 15, 2026
from 12:15 PM to 01:00 PM |
|---|---|
| Where | Bernstein Center Freiburg, Hansastr. 9a, Lecture Hall, Ground Floor |
| Contact Name | Martina Bacher |
| Add event to calendar |
 vCal iCal
|
ABSTRACT
Sparse recurrent spiking networks can generate rich collective dynamics, but the same instability that makes them expressive can also make their responses unreliable and their gradients ill-conditioned. I will discuss how Lyapunov spectra provide a common language for reliability, input control, and trainability in such systems. In balanced spiking networks, spike onset rapidness acts as a biophysical stability knob: increasing rapidness transforms conventional dense chaos into sparse, localized instability.
This sparse-chaos regime is especially susceptible to control by external input spike trains, which can suppress internal chaos and promote reliable state control by sensory drive. I will then briefly connect this dynamical-systems picture to learning in recurrent neural networks, where the same products of Jacobians that determine perturbation growth also underlie exploding and vanishing gradients during backpropagation through time. The talk will therefore trace a path from spiking-network chaos to input-driven reliability and learning dynamics.
About the speaker and his research