DIRA: Dynamic Domain Incremental Regularised Adaptation

TL;DR

DIRA is a novel method enabling deep neural networks to adapt to new operational domains during deployment using few samples, significantly reducing forgetting and improving robustness in dynamic environments.

Contribution

This paper introduces DIRA, a regularisation-based approach for online domain adaptation of DNNs that mitigates catastrophic forgetting with minimal samples.

Findings

DIRA outperforms existing methods on CIFAR-10C/100C benchmarks.

DIRA achieves state-of-the-art results in robustness to distribution shifts.

DIRA effectively balances adaptation and retention during online learning.

Abstract

Autonomous systems (AS) often use Deep Neural Network (DNN) classifiers to allow them to operate in complex, high-dimensional, non-linear, and dynamically changing environments. Due to the complexity of these environments, DNN classifiers may output misclassifications during operation when they face domains not identified during development. Removing a system from operation for retraining becomes impractical as the number of such AS increases. To increase AS reliability and overcome this limitation, DNN classifiers need to have the ability to adapt during operation when faced with different operational domains using a few samples (e.g. 2 to 100 samples). However, retraining DNNs on a few samples is known to cause catastrophic forgetting and poor generalisation. In this paper, we introduce Dynamic Incremental Regularised Adaptation (DIRA), an approach for dynamic operational domain adaption of DNNs using regularisation techniques. We show that DIRA improves on the problem of forgetting and achieves strong gains in performance when retraining using a few samples from the target domain. Our approach shows improvements on different image classification benchmarks aimed at evaluating robustness to distribution shifts (e.g.CIFAR-10C/100C, ImageNet-C), and produces state-of-the-art performance in comparison with other methods from the literature.