53 Making MetaDataCount Computer Vision News skin lesion models, in particular “trap sets”. Trap sets are artificially created biased sets with the training and test sets biased in opposite directions. The idea is that biased models “fall into the trap” and have poor performances on test data. Finally, as an example of debiasing solution, he presented their latest work: Test-Time Selection (TTS). The idea is to introduce a human-inthe-loop that will provide positive/negative keypoints in the image to encourage models to rely on robust features. As a closing remark, he suggested a promising approach of test-time debiasing to restrict the model's decision-making without altering its pretraining. If you want to stay updated about upcoming webinars and other events, sign up for our newsletter! As a final note, good luck to Hubert and Alceu in their upcoming PhD defenses :)
RkJQdWJsaXNoZXIy NTc3NzU=