Controllable Adaptive Learning

Lanyun Zhu, Tianrun Chen, Yan WANG

Keywords: Controllable Adaptive Learning

Abstract: As deep learning enabled unprecedented applications in versatile vision cognition tasks, researchers surged for the solutions of higher performance and more generalized algorithms, coming with expensive training and deployment to be applied in complex scenarios across domains. However, we argue that generalization and high performance are not always the ultimate goal in real-life with various applications and regulatory requirements. In this work, for the first time to our knowledge, we propose a Controllable Adaptive Learning (CAL) paradigm that allows the model to perform well on some data domains while performing poorly on others by control. We define the problem as a Controlled Multi-target Unsupervised Domain Adaptation (CMUDA) Task. Without the need to access labels in the target domain, we make the model perform poorly on certain target domains through a novel distribution different loss function design. We then introduced two easy-to-use control methods, namely implicit representation controller and explicit text-prompt controller, to regain access to the high-performance result with little effort, without the need to retrain the entire network. Extensive experiments demonstrated the effectiveness of our approach. We believe that our CAL paradigm will lead to an emerging trend for future research. Our code is at *URL*.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning