Stochastic memristive devices for low cost learning of spatiotemporal signals in spiking neural networks

Abstract Resistive switching devices are an excellent candidate for dedicated neural network hardware. They offer extremely low-power in-memory computing substrates for edge computing tasks like health monitoring. But, the imprecise and random conductance changes in these devices make deploying neural networks on such hardware significantly challenging. In this regard, biological random networks, known as liquid state machines (LSM), can be helpful. Using them as inspiration we can utilize the imprecise nature of the switching process for a low-cost training approach to learning in spiking recurrent neural networks. We rely on the inherent non-determinism associated with the conductance states in memristive devices to initialize the random weight matrices within a memristive LSM. We also utilize the randomness of the resistive states to introduce heterogeneity in the neuron parameters. The significance of the proposed approach is evaluated using arrhythmia and seizure detection edge computing tasks. For classification tasks using two datasets, our approach reduces the number of computational operations in the backward pass by factors of up to 66 × for the MIT-BIH arrhythmia dataset and 74 × for the CHB-MIT epileptic seizure dataset. The heterogeneity improves the network performance. We also show that our approach is resilient to write noise in memristive devices.