Task Arithmetic for Multi-Domain Visual Learning

Task Arithmetic is a novel approach that enables efficient fine-tuning of pre-trained models by manipulating weight spaces through arithmetic operations, such as addition and subtraction, without altering the original parameters. Although extensively explored in NLP, its application in multi-domain computer vision remains less explored. This study focuses on leveraging Task Arithmetic models for domain adaptation in image classification tasks. By modeling domain shifts through transformations in representation spaces, it enables robust generalization across domains while reducing reliance on large labeled datasets. Unlike conventional domain adaptation methods that require simultaneous access to source and target domain data, Task Arithmetic removes this dependency, providing a modular and scalable alternative. This work marks the first successful application of Task Arithmetic for visual domain adaptation, demonstrating its potential as a cost-effective and interpretable solution.