The Utility of Feature Reuse: Transfer Learning in Data-Starved Regimes

TL;DR

This paper investigates how transfer learning, especially feature reuse and overparameterization, improves CNN performance in data-starved regimes with fewer than 100 labeled samples, including generalization to OOD data.

Contribution

It demonstrates the effectiveness of transfer learning with overparameterized models and feature reuse in very limited data scenarios, supported by visual explanations.

Findings

Transfer learning improves CNN accuracy in data-starved regimes.

Overparameterization and feature reuse are key to successful transfer learning.

Transfer learning enhances generalization to out-of-distribution data.

Abstract

The use of transfer learning with deep neural networks has increasingly become widespread for deploying well-tested computer vision systems to newer domains, especially those with limited datasets. We describe a transfer learning use case for a domain with a data-starved regime, having fewer than 100 labeled target samples. We evaluate the effectiveness of convolutional feature extraction and fine-tuning of overparameterized models with respect to the size of target training data, as well as their generalization performance on data with covariate shift, or out-of-distribution (OOD) data. Our experiments demonstrate that both overparameterization and feature reuse contribute to the successful application of transfer learning in training image classifiers in data-starved regimes. We provide visual explanations to support our findings and conclude that transfer learning enhances the performance of CNN architectures in data-starved regimes.