Early Stopping Reduces Operations in Binary Neural Networks

Binary neural networks are highly appealing for deployment in resource-constrained environments due to their small footprint and low-power inference capabilities. While they convert dot products into sign-controlled additions or subtractions, the total number of operations typically remains the same, as each neuron still accumulates all its inputs before determining its binary output. This process can be wasteful, as often only the final sign matters. The core observation of this research is that during the accumulation process, the running partial sum frequently deviates so far from zero that its ultimate sign becomes highly predictable long before all input terms have been processed. Any subsequent contributions would change the sum's magnitude but not its final sign. This insight forms the basis of a new post-training early-stopping mechanism. The proposed method characterizes the behavior of running accumulations on the training dataset to predict the final sign as early as possible, without retraining any model parameters. When applied to VGG11 on the CIFAR-10 dataset, this technique eliminated 86.6% of the accumulation terms in the deepest convolution with only a 0.37-point accuracy drop. When used across the three deepest convolutions simultaneously, it reduced full-network arithmetic by 25% for a 1.36-point accuracy drop, significantly enhancing efficiency for constrained deployments.