A Parameter-Efficient Learning Framework with Memorized Facts

Conventional supervised learning approaches focus on direct mapping from input features to output labels. After training, these models predict test labels independently, without leveraging the training data and their underlying relationships. To better utilize both the training data and their associations, we propose Memory-Associated Differential (MAD) Learning, a novel learning paradigm. Our approach introduces a Memory component to store training data, learns label differences and feature associations through differential equations and sampling methods, and predicts unknown labels by combining memorized facts with learned differential patterns in a geometrically meaningful way. We first validate the framework on unary tasks (Weekday Prediction and Image Recognition), then extend it to binary link prediction, where it outperforms strong baselines.