Deep Without Gradients: A Forensic Analysis of Local Plasticity

Deep Without Gradients: A Forensic Analysis of Local Plasticity explores the limitations and potential of biologically plausible learning in Artificial Neural Networks by introducing the Event Gated Local Plasticity (EGLP) framework. EGLP combines Hebbian plasticity, Direct Feedback Alignment (DFA), and sparse surprise-triggered update gating to enable local hidden-layer learning without end-to-end gradient propagation. Through extensive experiments on MNIST and CIFAR-10, the study compares EGLP against supervised backpropagation, Hebbian learning, random frozen features, and hybrid pretraining approaches. The results demonstrate that EGLP can learn meaningful representations while significantly reducing computational cost and hidden-layer update frequency, achieving nearly 3× lower FLOPs with sparse update events. However, the work also highlights critical challenges of strictly local plasticity mechanisms, particularly in deep architectures where credit assignment and representation learning become increasingly difficult. The study provides an in-depth analysis of scaling behavior, efficiency trade-offs, and representational limitations, offering insights into the future of biologically inspired and neuromorphic learning systems.