Jonathan Frankle & Michael Carbin (2018)
arXiv.
DOI: https://doi.org/10.48550/arxiv.1803.03635
Abstract. Proposes the Lottery Ticket Hypothesis: dense, randomly initialised feedforward networks contain sparse subnetworks ('winning tickets') that, trained in isolation, can match the accuracy of the original network. Provides theoretical motivation for pruning.
Tags: efficiency pruning lottery-ticket
Cited in: