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

Area of Science:

• Materials Science
• Neuroscience
• Computer Engineering

Background:

• Training spiking neural networks (SNNs) in neuromorphic hardware is challenging due to non-differentiable activations and limited energy-efficient platforms.
• Existing methods struggle with the computational demands and hardware constraints of SNNs.

Purpose of the Study:

• To develop a flexible, trainable neuromorphic hardware platform for efficient SNN training.
• To implement surrogate gradient computation for backpropagation in SNNs using novel organic transistors.

Main Methods:

• Fabrication of tunable textile-based vertical organic electrochemical transistors (TT-vOECTs) using a solvent interdiffusion solidification spinning strategy.
• Development of a conditionally activated backpropagation (CAB) mechanism using dual TT-vOECTs for event-driven synaptic updates.
• Integration of the CAB strategy into a convolutional SNN for signal classification.

Main Results:

• The TT-vOECT exhibits nonlinear transfer characteristics approximating a Sigmoid derivative, suitable for surrogate gradients.
• The device-driven CAB strategy enabled sparse, event-driven weight updates, reducing computational overhead.
• High-accuracy classification of electroencephalogram signals for neurological disorder diagnosis was achieved, comparable to software baselines with ≈20% reduced computational redundancy.

Conclusions:

• A scalable and flexible organic hardware platform for trainable neuromorphic systems was established.
• The TT-vOECT and CAB mechanism offer a pathway towards biologically inspired learning dynamics in neuromorphic computing.
• This work demonstrates the potential of organic electronics for advanced AI applications.