Order parameters and phase transitions of continual learning in deep neural networks

TL;DR

This paper develops a statistical-mechanics theory for continual learning in deep neural networks, identifying key order parameters that predict forgetting and interference, and suggests architectural strategies to mitigate these issues.

Contribution

It introduces a theoretical framework with order parameters that characterize task relations and network architecture effects on continual learning performance.

Findings

Increasing network depth reduces task interference.

Task similarity predicts phase transitions in learning performance.

Low similarity can cause catastrophic forgetting of new tasks.

Abstract

Continual learning (CL) enables animals to learn new tasks without erasing prior knowledge. CL in artificial neural networks (NNs) is challenging due to catastrophic forgetting, where new learning degrades performance on older tasks. While various techniques exist to mitigate forgetting, theoretical insights into when and why CL fails in NNs are lacking. Here, we present a statistical-mechanics theory of CL in deep, wide NNs, which characterizes the network's input-output mapping as it learns a sequence of tasks. It gives rise to order parameters (OPs) that capture how task relations and network architecture influence forgetting and anterograde interference, as verified by numerical evaluations. For networks with a shared readout for all tasks (single-head CL), the relevant-feature and rule similarity between tasks, respectively measured by two OPs, are sufficient to predict a wide range of CL behaviors. In addition, the theory predicts that increasing the network depth can effectively reduce interference between tasks, thereby lowering forgetting. For networks with task-specific readouts (multi-head CL), the theory identifies a phase transition where CL performance shifts dramatically as tasks become less similar, as measured by another task-similarity OP. While forgetting is relatively mild compared to single-head CL across all tasks, sufficiently low similarity leads to catastrophic anterograde interference, where the network retains old tasks perfectly but completely fails to generalize new learning. Our results delineate important factors affecting CL performance and suggest strategies for mitigating forgetting.