Next Shift

A digit recogniser trained on clean scans meets a new environment: grey, low-contrast, textured, rotated and blurred images. Label a few target images: the model fine-tunes its top layers on your labels in the browser, and its predictions on held-out target images update.

Target environment is syntheticLive in-browser adaptationUCI digits (CC BY 4.0)Model repo

Label target images

Held-out target accuracy

0your labels
–Next Shift
–frozen source features + linear head

Loading…

Next Shift predictions on held-out target images

Recorded benchmark

Accuracy vs labeled examples