Abstract
Memristors, resistors with memory whose outputs depend on the history oftheir inputs, have been used with success in neuromorphic architectures,particularly as synapses or non-volatile memories. A neural network based onmemristors could show advantages in terms of energy conservation and open uppossibilities for other learning systems to be adapted to a memristor-basedparadigm, both in the classical and quantum learning realms. No model for sucha network has been proposed so far. Therefore, in order to fill this gap, weintroduce models for single and multilayer perceptrons based on memristors. Weadapt the delta rule to the memristor-based single-layer perceptron and thebackpropagation algorithm to the memristor-based multilayer perceptron. We ransimulations of both the models and the training algorithms. These showed thatboth of them perform well and in accordance with Minsky-Papert's theorem, whichmotivates the possibility of building memristor-based hardware for a physicalneural network.