Contrastive Learning in Memristor-based Neuromorphic Systems
Cory Merkel, Alexander Ororbia
TL;DR
The paper tackles the challenge of training neuromorphic systems without backpropagation by implementing a hardware-friendly variant of contrastive learning called CSDP, derived from the Forward-Forward framework. It presents a 45 nm CMOS+memristor prototype that uses LIF neurons, trace circuits, and dual-memristor synapses to perform contrastive learning on spike trains. The results demonstrate that the hardware can learn a simple logic function without gradient-based multi-layer backpropagation, offering a promising route toward energy-efficient on-chip training. The work provides detailed hardware building blocks and a near proof-of-concept for CSDP in memristor-based neuromorphic platforms, with future work to compare against STDP, feedback alignment, and surrogate BP on more complex tasks.
Abstract
Spiking neural networks, the third generation of artificial neural networks, have become an important family of neuron-based models that sidestep many of the key limitations facing modern-day backpropagation-trained deep networks, including their high energy inefficiency and long-criticized biological implausibility. In this work, we design and investigate a proof-of-concept instantiation of contrastive-signal-dependent plasticity (CSDP), a neuromorphic form of forward-forward-based, backpropagation-free learning. Our experimental simulations demonstrate that a hardware implementation of CSDP is capable of learning simple logic functions without the need to resort to complex gradient calculations.