Entropy Minimization
Share of the total weight, by predicted class (10-batch average)
Weight of one sample, by predicted class
Every sample has a weight of 1, whatever its class, so each class’s share of the weight simply follows the predictions.
NeurIPS 2026 Spotlight Paper Main Track Top ~3.7%
Entropy Minimization without Model Collapse: Mitigating Prediction Bias in Medical Imaging
*Shared first authorship · #Shared last authorship
Under distribution shift, feature clusters merge and the model develops a prediction bias: it over-predicts one class although the true classes are balanced. Entropy Minimization weights every sample equally and amplifies this bias until model collapse; our DSBR reweights samples by their predicted class and dampens it.
Running live in your browser
Share of the total weight, by predicted class (10-batch average)
Weight of one sample, by predicted class
Every sample has a weight of 1, whatever its class, so each class’s share of the weight simply follows the predictions.
Share of the total weight, by predicted class (10-batch average)
Weight of one sample, by predicted class
A tall bar marks a suppressed class, one the model currently predicts rarely. DSBR raises its weight so that it is not drowned out.
Both methods adapt the same small synthetic model on identical batches, using the update rules from the paper’s code. We report model collapse when one class receives at least 95% of the predictions on held-out samples. DSBR uses a decay of α = 0.9.