377-Jaxley: Simulation for Biophysical Neural Models

Paper Talk · Jan 7, 2026 · 20m

The paper introduces Jaxley, a novel differentiable simulation framework designed to bridge the gap between detailed biophysical neuron models and modern machine learning techniques. By utilizing JAX-based automatic differentiation and GPU acceleration, Jaxley allows researchers to optimize complex neural parameters—such as ion channel conductances and synaptic weights—using gradient descent rather than slower, gradient-free methods. The authors demonstrate that this tool can efficiently fit models to physiological recordings, train recurrent networks for memory tasks, and even scale to…