Abstract
Memristors are resistive circuit elements that allow changes in resistance using voltage pulses. The neuromorphic computing group at Boise State University has been building memristors and using them in neural networks. Learning in the network is done in an unsupervised manner using spike timing dependent plasticity (STDP). Unsupervised learning occurs without explicit signals controlling the neuron's learning and without any feedback from the system. Through STDP, the neuron becomes sensitive to repeating spike patterns in a spike train. The synaptic weights are increased on early firing afferents and the postsynaptic latencies are decreased. We are currently evaluating how the learning algorithm works. Simulations in Matlab focusing on unsupervised learning using STDP will provide a good application for memristors. Since STDP might be the mechanism through which the synapses in the brain operate, modeling it using memristors is a good choice for neural networks.
The final report is available here.