Alleviating Representational Shift for Continual Fine-tuning

TL;DR

This paper introduces ConFiT, a novel fine-tuning approach that mitigates representational shift in continual learning by using cross-convolution batch normalization and hierarchical fine-tuning, leading to improved performance.

Contribution

The paper proposes ConFiT, a new fine-tuning method that addresses both feature and intermediate layer shifts, enhancing continual learning effectiveness.

Findings

Outperforms state-of-the-art methods on four datasets.

Reduces catastrophic forgetting with lower storage overhead.

Effectively maintains feature representations during fine-tuning.

Abstract

We study a practical setting of continual learning: fine-tuning on a pre-trained model continually. Previous work has found that, when training on new tasks, the features (penultimate layer representations) of previous data will change, called representational shift. Besides the shift of features, we reveal that the intermediate layers' representational shift (IRS) also matters since it disrupts batch normalization, which is another crucial cause of catastrophic forgetting. Motivated by this, we propose ConFiT, a fine-tuning method incorporating two components, cross-convolution batch normalization (Xconv BN) and hierarchical fine-tuning. Xconv BN maintains pre-convolution running means instead of post-convolution, and recovers post-convolution ones before testing, which corrects the inaccurate estimates of means under IRS. Hierarchical fine-tuning leverages a multi-stage strategy to fine-tune the pre-trained network, preventing massive changes in Conv layers and thus alleviating IRS. Experimental results on four datasets show that our method remarkably outperforms several state-of-the-art methods with lower storage overhead.