Not Every Subject Should Stay: Machine Unlearning for Noisy Engagement Recognition
The study tested if removing problematic subjects from a pre-trained engagement recognition model could recover significant performance gains without full retraining. By identifying subjects with high training loss, researchers applied a lightweight unlearning update. Results showed this method recovered 89.3% to 92.5% of the benefit seen when retraining only on kept data, proving it's a practical way to sanitize models after initial training.