Not Every Subject Should Stay: Machine Unlearning for Noisy Engagement Recognition
Researchers have developed a machine unlearning technique to remove the influence of specific subjects from trained models without requiring a full retraining process. This method, applied to engagement recognition datasets like DAiSEE and EngageNet, aims to sanitize models by identifying and excluding problematic data subsets. The unlearned models achieved significant performance recovery, nearing that of models retrained from scratch, at a fraction of the computational cost. AI
IMPACT Offers a more efficient method for correcting biases in trained models, potentially reducing costs associated with data curation and retraining.