Abstract
This work addresses how to naturally adopt the l2-norm cosine similarity in the neuromemristive system and studies the unsupervised learning performance on handwritten digit image recognition. Proposed architecture is a two-layer fully connected neural network with a hard winner-take-all (WTA) learning module. For input layer, we propose single-spike temporal code that transforms input stimuli into the set of single spikes with different latencies and voltage levels. For a synapse model, we employ a compound memristor where stochastically switching binary-state memristors connected in parallel, which offers a reliable and scalable multi-state solution for synaptic weight storage. Hardware-friendly synaptic adaptation mechanism is proposed to realize spike-timing-dependent plasticity learning. Input spikes are sent out through those memristive synapses to each and every integrate-and-fire neuron in the fully connected output layer, where the hard WTA network motif introduces the competition based on cosine similarity for the given input stimuli. Finally, we present 92.64% accuracy performance on unsupervised digit recognition with only single-epoch MNIST dataset training via high-level simulations, including extensive analysis on the impact of system parameters.
Original language | English |
---|---|
Article number | 38 |
Journal | ACM Journal on Emerging Technologies in Computing Systems |
Volume | 18 |
Issue number | 2 |
DOIs | |
State | Published - Apr 2022 |
Keywords
- MNIST
- Unsupervised learning
- compound memristor
- digit recognition
- neuromorphic computing
- single-spike temporal code
- spiking neural network
- winner-take-all