Confidence and Dispersity Speak: Characterising Prediction Matrix for Unsupervised Accuracy Estimation

Weijian Deng, Yumin Suh, Stephen Gould, Liang Zheng

Keywords: Out-of-distribution generalization, Unsupervised Accuracy Estimation, Prediction DIversity, Distribution Shift

TL;DR: This work proposes a simple but effective method (prediction diversity) to predict how well a model generalize to out-of-distribution datasets

Abstract: This work focuses on estimating how well a model performs on out-of-distribution (OOD) datasets without using labels. Our intuition is that a well-performing model should give predictions with high confidence and high dispersity. While recent methods study the prediction confidence, this work newly finds dispersity is another informative cue. Confidence reflects whether the individual prediction is certain; dispersity indicates how the overall predictions are distributed across all categories. To achieve a more accurate estimation, we propose to jointly consider these two properties by using the nuclear norm of the prediction matrix. In our experiments, we extensively validate the effectiveness of nuclear norm for various models (e.g., ViT and ConvNeXt), different datasets (e.g., ImageNet and CUB-200), and diverse types of distribution shifts (e.g., style shift and reproduction shift). We show that the nuclear norm is more accurate and robust in predicting OOD accuracy than existing methods. Lastly, we study the limitation of the nuclear norm and discuss potential directions.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning