EWC

Elastic Weight Consolidation (EWC) is a continual learning method designed to mitigate catastrophic forgetting when neural networks learn tasks sequentially. It protects weights that are important to previously learned tasks while allowing less essential weights to adapt to new ones.

It does so by adding a regularization term to the loss for the current task. After learning a task, the method estimates the importance of each parameter using the Fisher Information Matrix. The diagonal elements F_i describe how sensitive performance is to changes in theta_i. The objective becomes L_total = L_current + (lambda/2) sum_i F_i (theta_i - theta_i*)^2, where theta_i* is the old value and lambda controls the penalty strength.

Consequences: Important weights are protected, reducing forgetting while allowing less important weights to adapt. The approach relies on a diagonal FIM approximation for efficiency and may be less effective if tasks are highly dissimilar or many tasks are learned. Variants include Online EWC and methods that relax the diagonal assumption or combine EWC with replay.

Applications and context: EWC has been used in image classification, reinforcement learning, and robotics. It is part of a family of regularization-based continual learning methods, alongside Synaptic Intelligence and Memory Aware Synapses, and contrasts with replay-based approaches that reuse past data.