Selective Plasticity in Spiking Neural Networks for Energy-Efficient Unsupervised On-Chip Learning: Toward Self-Adapting Implantable Neuromorphic Biosignal Processors
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Self-adapting implantable biosignal processors motivate energy-efficient on-chip learning, but this thesis evaluates the underlying SNN methods on MNIST rather than patient data. An STDP network is decomposed into synaptic weights, adaptive thresholds, and firing-rate statistics; freeze ablations identify their timescales and motivate confidence- and neuron-selectivity-gated updates. Per-neuron selectivity and maturity are then used to diagnose severe class imbalance, derive four local ratio-agnostic plasticity mechanisms, and organize neurons topographically for window-restricted inference. On controlled rare-event MNIST tasks, selective adaptation reduces weight-commit traffic by approximately 80% during adaptation and 92% at steady state. At 1:1000 imbalance, sensitivity rises from 41% to 72%; windowing retains 68% sensitivity while reducing modeled energy from 6.7 to 1.8 μJ. A fixed-point FPGA baseline with on-chip STDP reaches 86.2% MNIST accuracy, 131 μs latency, and 22 μJ measured programmable-logic dynamic energy per inference. These results support selective plasticity toward implantable neuromorphic learning systems.