Improving Robustness In Analog Neural Networks By Introducing Sparse Connectivity

Sparsity in the connectivity of an analog neural network can both reduce the amount of energy needed for multiply-accumulate operations as well as mitigate the effects of noise and random variations within the network. This paper gives an example for how sparsity can decrease the uncertainty in noisy, low-precision networks and compares common techniques for generating sparse network architectures. This work also presents a better method for generating small sparse neural networks.