Image Completion using a Sparse Probabilistic Spin Logic Network

Abstract

Probabilistic Spin Logic (PSL) is a recently proposed computational paradigm implemented using unstable stochastic units called probabilistic bits, or pbits. The stochastic nature of the pbits is responsible for its unique invertibility property. In this paper, we have exploited the invertibility property to complete a partial image. The digit image set consists of ten images of digits 0 to 9 of size $5 \times 3$ pixels for training the PSL network. Results show that the fully connected PSL (FC-PSL) network successfully recovers the unclamped pixels for all the digits. Moreover, we propose an area-efficient sparsely connected PSL (SC-PSL) network with ~ 42% of the original connections. First, we perform the sparsification on an FC-PSL network using the conventional weight pruning methods. Next, we propose a method to predict the SC-PSL network structure using mutual information based on the perspective of information theory.