LWR-Net: Learning without retraining for scalable multi-task adaptation and domain-agnostic generalisation

Abstract

In recent years, deep learning-based multi-class and multi-task classification have gained significant attention across various domains of computer vision. However, current approaches often struggle to incorporate new classes efficiently due to the computational burden of retraining large neural networks from scratch. This limitation poses a significant obstacle to the deployment of deep learning models in real-world intelligent systems. Although continual learning has been proposed to overcome this challenge, it remains constrained by catastrophic forgetting. To address these limitations, this study introduces a new framework called Learning Without Retraining (LWR-Net), developed for multi-class and multi-task adaptation, allowing networks to adapt to new classes with minimal training requirements. Specifically, LWR-Net incorporates four key components: (i) task-guided self-supervised learning with a dual-attention mechanism to enhance feature generalisation and selection; (ii) task-based model fusion to improve feature representation and generalisation; (iii) multi-task learning to generalise classifiers across diverse tasks; and (iv) decision fusion of multiple classifiers to improve overall performance and reduce the likelihood of misclassification. LWR-Net was evaluated across diverse tasks to demonstrate its effectiveness in integrating new data, classes, or tasks. These include: (i) a medical case study detecting abnormalities in five distinct bone structures; (ii) a surveillance case study detecting violence in three different settings; and (iii) a geology case study identifying lateral changes in soil compaction using ground-penetrating radar across two datasets. The results show that LWR-Net achieves state-of-the-art performance across all three scenarios, successfully accommodates new learning objectives while preserving performance, eliminating the need for complete retraining cycles. Moreover, the use of gradient-weighted class activation mapping (Grad-CAM) confirmed that the models focused on relevant regions of interest. LWR-Net offers several benefits, including improved generalisation, enhanced performance, and the capacity to train on new data without catastrophic failures. The source code is publicly available at: https://github.com/LaithAlzubaidi/Learning-to-Adapt.