On-Chip Learning with a 15-neuron Digital Oscillatory Neural Network Implemented on ZYNQ Processor
Abstract
Real-time on-chip learning is an important feature for current neuromorphic computing to enable smart embedded systems capable of learning. Neuromorphic computing based on Oscillatory Neural Networks (ONNs) are networks of coupled oscillators computing with phase information. ONNs with fully-connected connections can perform autoassociative memory applications when trained with unsupervised learning rules. In this paper, we propose for the first time an architecture to perform on-chip learning with a digitally implemented ONN. We implement the digital ONN with programmable logic of a ZYNQ processor and we perform learning on the processing system of the same chip. We validate our solution on a 15-neuron ONN trained with either Hebbian or Storkey learning rules up to three patterns. We report a stable resource utilization for both learning rules and timing from 119 µs (Hebbian) to 163 µs (Storkey). Additionally, accuracy is equal to the off-chip learning implementation.
Origin | Files produced by the author(s) |
---|