Zero-shot generalization across architectures for visual classification
Computer Vision and Pattern Recognition
2024-05-06 v4 Artificial Intelligence
Machine Learning
Abstract
Generalization to unseen data is a key desideratum for deep networks, but its relation to classification accuracy is unclear. Using a minimalist vision dataset and a measure of generalizability, we show that popular networks, from deep convolutional networks (CNNs) to transformers, vary in their power to extrapolate to unseen classes both across layers and across architectures. Accuracy is not a good predictor of generalizability, and generalization varies non-monotonically with layer depth.
Cite
@article{arxiv.2402.14095,
title = {Zero-shot generalization across architectures for visual classification},
author = {Evan Gerritz and Luciano Dyballa and Steven W. Zucker},
journal= {arXiv preprint arXiv:2402.14095},
year = {2024}
}
Comments
Accepted as a Tiny Paper at ICLR 2024. Code available at https://github.com/dyballa/generalization/tree/ICLR2024TinyPaper