The Entropy Enigma: Success and Failure of Entropy Minimization

This repository contains the code used in our ICML 2024 paper. Our method allows you to estimate the accuracy of a model on a given dataset, while the model adapts to the dataset.

We show that method that use entropy minimization to adapt to their inputs work by clustering the embeddings of their inputs. In essence, this works great as long as the clusters formed by entropy minimization stay close to the embeddings of training images. Eventually, entropy minimization causes the model to embed input images far away from the embeddings of training images. This is what eventually decreases the model's accuracy to nearly 0, as we showed in our paper RDumb: A simple approach that questions our progress in continual test-time adaptation.

Using this, we propose a method to estimate the accuracy of a classifier on a given input dataset, without having access to the labels. Our method works by looking at how the model predictions change over time, as the model uses entropy minimization to adapt to its inputs. Concretely, we take a set of 1,000 input images, and run them through the model before adaptation. We record the model's predicted class for every image. Then, at the end of adaptation, we run the same set of images through the model, and compare the model's current prediction to its prediction at the start. If the prediction changed, we call that a label flip. To estimate accuracy, we plug the total number of label flips into our flips-to-accuracy function f.

Intuitively, if a model's accuracy is high on a certain dataset, embeddings of samples will probably be well clustered at the start, and not change much during entropy minimization. This would yield a low number of label flips. Conversely, if a model's accuracy is low, the embedding space will be disorganized, and entropy minimization will move the embeddings around, yielding many label flips. We elaborate on this in our paper.

Getting started

Running estimation requires just one argument. If you want to additionally use temperature scaling, you have to also provide a calibration dataset.

python3 main.py \
    --data /path/to/dset/ \            # dataset you want to estimate
    --valdata /path/to/val_dset \      # validation dataset, if you want a temperature scaled model

BibTeX

If you find this work helpful, please use the following citation:

@inproceedings{pres2024entropy,
  title={The Entropy Enigma: Success and Failure of Entropy Minimization},
  author={Press, Ori and Shwartz-Ziv, Ravid and LeCun, Yann and Bethge, Matthias},
  booktitle={International Conference on Machine Learning},
  year={2024},
  organization={PMLR}
}

Acknowledgements

Much of model code is based on RDumb which is based on EATA code. We also use Temperature Scaling. See the LICENSE for more details and licenses.