The Pitfalls of Memorization: When Memorization Hurts Generalization
This addresses a fundamental issue in machine learning where memorization leads to poor generalization, particularly under distribution shifts, but the approach is incremental as it builds on existing training methods.
The paper tackles the problem of neural networks memorizing exceptions and relying on spurious correlations, which harms generalization, and proposes memorization-aware training (MAT) to improve generalization under distribution shifts.
Neural networks often learn simple explanations that fit the majority of the data while memorizing exceptions that deviate from these explanations.This behavior leads to poor generalization when the learned explanations rely on spurious correlations. In this work, we formalize the interplay between memorization and generalization, showing that spurious correlations would particularly lead to poor generalization when are combined with memorization. Memorization can reduce training loss to zero, leaving no incentive to learn robust, generalizable patterns. To address this, we propose memorization-aware training (MAT), which uses held-out predictions as a signal of memorization to shift a model's logits. MAT encourages learning robust patterns invariant across distributions, improving generalization under distribution shifts.