Shayan Shahrabi-Farahani · Shahid Beheshti University
The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks
Abstract
Neural networks often contain millions of parameters. But do they really need all of them? This paper argues that, hidden inside a randomly initialized network, there may already exist a small winning ticket capable of reaching the same accuracy as the full model when trained in isolation. It's a simple idea with far-reaching implications for pruning, efficiency, and our understanding of why deep learning works.
