Convergence and Implicit Bias of Gradient Descent on Continual Linear Classification
This provides theoretical insights into implicit bias and forgetting in continual learning, which is incremental but addresses a key challenge for AI systems that learn sequentially.
The paper tackles the problem of catastrophic forgetting in continual learning by analyzing gradient descent on multiple linear classification tasks, showing that the classifier converges to the joint max-margin solution when tasks are jointly separable and presented cyclically or randomly, with forgetting vanishing to zero as cycles repeat.
We study continual learning on multiple linear classification tasks by sequentially running gradient descent (GD) for a fixed budget of iterations per task. When all tasks are jointly linearly separable and are presented in a cyclic/random order, we show the directional convergence of the trained linear classifier to the joint (offline) max-margin solution. This is surprising because GD training on a single task is implicitly biased towards the individual max-margin solution for the task, and the direction of the joint max-margin solution can be largely different from these individual solutions. Additionally, when tasks are given in a cyclic order, we present a non-asymptotic analysis on cycle-averaged forgetting, revealing that (1) alignment between tasks is indeed closely tied to catastrophic forgetting and backward knowledge transfer and (2) the amount of forgetting vanishes to zero as the cycle repeats. Lastly, we analyze the case where the tasks are no longer jointly separable and show that the model trained in a cyclic order converges to the unique minimum of the joint loss function.