Fiber Transistors as a Hardware Surrogate Gradient for Backpropagation in Spiking Neural Networks.

Abstract

Efficient training of spiking neural networks (SNNs) in flexible neuromorphic hardware remains a major challenge due to the non-differentiable nature of spiking activations and the limited availability of trainable, energy-efficient device platforms. Here, a tunable textile-based vertical organic electrochemical transistor (TT-vOECT) enables implementation of surrogate gradients computation for backpropagation in SNNs. Using a solvent interdiffusion solidification spinning strategy, Plateau-Rayleigh instabilities are overcome to fabricate coaxial trilayer fibers with well-defined heterojunctions between PEDOT:PSS and BBL. The resulting TT-vOECT exhibits nonlinear transfer characteristics that closely approximate the Sigmoid derivative. A conditionally activated backpropagation (CAB) mechanism is further proposed, in which synaptic updates are gated by both surrogate gradient magnitude and input spike, is implemented using reconfigurable logic arrays based on dual TT-vOECTs. This framework enables sparse, event-driven weight updates with reduced computational overhead. Integrated into a convolutional SNN, the device-driven CAB strategy achieves high-accuracy classification of electroencephalogram signals for multiclass neurological disorder diagnosis, comparable to software baselines, while reducing computational redundancy by ≈20%. The results establish a scalable and flexible organic hardware platform for trainable neuromorphic systems with biologically inspired learning dynamics.