{Michael Carbin Jonathan Frankle}

Abstract

Neural network pruning techniques can reduce the parameter counts of trainednetworks by over 90%, decreasing storage requirements and improvingcomputational performance of inference without compromising accuracy. However,contemporary experience is that the sparse architectures produced by pruningare difficult to train from the start, which would similarly improve trainingperformance. We find that a standard pruning technique naturally uncovers subnetworkswhose initializations made them capable of training effectively. Based on theseresults, we articulate the "lottery ticket hypothesis:" dense,randomly-initialized, feed-forward networks contain subnetworks ("winningtickets") that - when trained in isolation - reach test accuracy comparable tothe original network in a similar number of iterations. The winning tickets wefind have won the initialization lottery: their connections have initialweights that make training particularly effective. We present an algorithm to identify winning tickets and a series ofexperiments that support the lottery ticket hypothesis and the importance ofthese fortuitous initializations. We consistently find winning tickets that areless than 10-20% of the size of several fully-connected and convolutionalfeed-forward architectures for MNIST and CIFAR10. Above this size, the winningtickets that we find learn faster than the original network and reach highertest accuracy.