On-Chip Incremental Learning Based on Unsupervised STDP Implementation

Spiking neural network (SNN), a bio-inspired neuron network, utilizes a learning rule named spike-timing-dependent plasticity (STDP) to achieve high-performance unsupervised learning. However, it may suffer from catastrophic forgetting when the distribution of new data significantly differs from that of old data. To address this issue, an incremental learning implementation for ship classification is presented in this paper. We develop an incremental learning algorithm based on STDP and corresponding platform. A competitive SNN is built into our algorithm, and add-STDP is utilized to update the weights of network for efficient learning. To enhance learning performance, we incorporate weight decay. And to avoid catastrophic forgetting, we incorporate data replay. The corresponding learning platform consists of the FPGA Zynq 7100 and the STDP neuromorphic prototype chip, and our algorithm is executed on the chip. We evaluate the ship classification task on our platform, which demonstrates the superior potential of our on-chip implementation for incremental learning.