The Lottery Ticket Hypothesis: Finding Sparse,…

The Lottery Ticket Hypothesis asks why a small neural network can be difficult to train even when pruning proves that it has enough capacity. Read it to see how iterative magnitude pruning discovers a sparse mask, why resetting its surviving weights to their exact original values matters, and why the resulting ticket is still expensive to find and awkward for ordinary hardware.

Reading focus: Why a winning ticket is the combination of a sparse mask and the surviving weights' original initialization, not the sparse topology alone. How train, prune, reset, and repeat turns post-training magnitude into a retrospective search signal for a trainable subnetwork. What the MNIST and CIFAR-10 results show, and how repeated training, unstructured sparsity, learning-rate sensitivity, and untested generality limit the claim.

ICLR 2019. Jonathan Frankle and Michael Carbin. 30 min read, easy difficulty.