One ticket to win them all: generalizing lottery ticket initializations across datasets and optimizers

TL;DR

This paper investigates whether lottery ticket initializations can be transferred across datasets and optimizers, finding that larger datasets produce more generalizable tickets that perform well across various configurations.

Contribution

The study demonstrates that winning ticket initializations can generalize across datasets and optimizers, especially when generated from larger datasets, indicating broader applicability.

Findings

Winning tickets transfer well across datasets in the natural images domain.

Larger datasets produce more transferable winning tickets.

Winning tickets also generalize across different optimizers.

Abstract

The success of lottery ticket initializations (Frankle and Carbin, 2019) suggests that small, sparsified networks can be trained so long as the network is initialized appropriately. Unfortunately, finding these "winning ticket" initializations is computationally expensive. One potential solution is to reuse the same winning tickets across a variety of datasets and optimizers. However, the generality of winning ticket initializations remains unclear. Here, we attempt to answer this question by generating winning tickets for one training configuration (optimizer and dataset) and evaluating their performance on another configuration. Perhaps surprisingly, we found that, within the natural images domain, winning ticket initializations generalized across a variety of datasets, including Fashion MNIST, SVHN, CIFAR-10/100, ImageNet, and Places365, often achieving performance close to that of winning tickets generated on the same dataset. Moreover, winning tickets generated using larger datasets consistently transferred better than those generated using smaller datasets. We also found that winning ticket initializations generalize across optimizers with high performance. These results suggest that winning ticket initializations generated by sufficiently large datasets contain inductive biases generic to neural networks more broadly which improve training across many settings and provide hope for the development of better initialization methods.