Entropy Minimization
An unsupervised learning method that encourages a model to produce high-confidence predictions by minimizing the Shannon entropy of its predictive distribution. It is commonly used in test-time adaptation, where a model trained on a source domain is adjusted on unlabeled target data. By reducing prediction uncertainty, the method promotes sharper output distributions and more decisive class assignments, and it is often combined with other self-supervised or consistency-based objectives.
Timeline (last 90 days)
Mainstream entropy-based TTA methods rely on deterministic confidence and may reinforce overconfident errors under distribution shift.
Entropy minimization drives predictions toward one-hot by minimizing the Shannon entropy of the prediction distribution, but it does not test whether the predictions are correct.
Entropy is sensitive to temperature scaling and logit scale, and a model may lower entropy by amplifying logit magnitudes rather than improving feature representations.
All Facts (3)
Mainstream entropy-based TTA methods rely on deterministic confidence and may reinforce overconfident errors under distribution shift.
50%UnverifiedEntropy minimization drives predictions toward one-hot by minimizing the Shannon entropy of the prediction distribution, but it does not test whether the predictions are correct.
50%UnverifiedEntropy is sensitive to temperature scaling and logit scale, and a model may lower entropy by amplifying logit magnitudes rather than improving feature representations.
50%