An Empirical Analysis of Forgetting in Pre-trained Models with Incremental Low-Rank Updates

TL;DR

This paper investigates how low-rank adaptation (LoRA) impacts forgetting in pretrained models during incremental learning, revealing that LoRA rank influences forgetting and introduces a novel 'contextual' forgetting behavior in vision transformers.

Contribution

It studies the effect of LoRA rank on model forgetting and introduces the concept of 'contextual' forgetting in vision transformers, a novel observation in continual learning.

Findings

LoRA rank significantly affects forgetting of pretraining and downstream tasks.

Vision transformers exhibit 'contextual' forgetting not seen in residual networks.

LoRA merging impacts model plasticity and retention during incremental updates.

Abstract

Broad, open source availability of large pretrained foundation models on the internet through platforms such as HuggingFace has taken the world of practical deep learning by storm. A classical pipeline for neural network training now typically consists of finetuning these pretrained network on a small target dataset instead of training from scratch. In the case of large models this can be done even on modest hardware using a low rank training technique known as Low-Rank Adaptation (LoRA). While Low Rank training has already been studied in the continual learning setting, existing works often consider storing the learned adapter along with the existing model but rarely attempt to modify the weights of the pretrained model by merging the LoRA with the existing weights after finishing the training of each task. In this article we investigate this setting and study the impact of LoRA rank on the forgetting of the pretraining foundation task and on the plasticity and forgetting of subsequent ones. We observe that this rank has an important impact on forgetting of both the pretraining and downstream tasks. We also observe that vision transformers finetuned in that way exhibit a sort of ``contextual'' forgetting, a behaviour that we do not observe for residual networks and that we believe has not been observed yet in previous continual learning works.