The Lottery Ticket Hypothesis: Why Pruned Models Can Sometimes Learn Just as Well as Full Networks

The Lottery Ticket Hypothesis (LTH) proposes that within large neural networks exist smaller subnetworks—or “winning tickets”—that, when properly initialized and trained, can match or even outperform their full-sized counterparts. This article surveys key research exploring LTH’s methods, extensions, and limitations, including its applications in vision, NLP, and reinforcement learning. It examines how pruning, initialization, and iterative retraining contribute to model efficiency, generalization, and theoretical understanding, while also addressing open challenges around scalability, domain transfer, and the deeper mechanisms behind why LTH works.