eXplaining to Learn (eX2L): Regularization Using Contrastive Visual Explanation Pairs for Distribution Shifts
Researchers have introduced eXplaining to Learn (eX2L), a novel framework designed to improve model performance and interpretability when faced with distribution shifts. This method works by decoupling confounding features from a classifier's latent representations during training. eX2L achieves this by penalizing the similarity between activation maps from a primary classifier and those from a concurrently trained confounder classifier. The framework demonstrated significant improvements on the Spawrious Many-to-Many Hard Challenge benchmark, outperforming the current state-of-the-art. AI
IMPACT Introduces a new method for improving model robustness against distribution shifts, potentially enhancing reliability in real-world applications.