Divide and Conform: Unleashing Spatial Filter Atoms for Unsupervised Target Transferability

Gaurav Patel, Qiang Qiu

Keywords: Filter Decomposition, Domain Transferability, Efficiency

TL;DR: The paper presents Divide and Conform, a method parameter-efficient and interpretable framework for knowledge transferability of pre-trained ConvNets by selectively fine-tuning the spatial filter atoms.

Abstract: The straightforward fine-tuning of the pre-trained model for the target task, bears the risk of under-utilizing the foundational knowledge accrued by the pre-trained model, resulting in the sub-optimal utilization of transferable knowledge, consequently impeding peak performance on the target task. To address this, we introduce $\textit{Divide and Conform}$, aimed at augmenting the transferability of pre-trained convolutional neural networks (ConvNets), $\textit{in the absence of base data}$. This strategy exploits the mathematical equivalence of the convolution operation, conceptualizing it as a two-step process involving spatial-only convolution and channel combination. To achieve this, we decompose ($\textit{Divide}$) the filters of pre-trained ConvNets into spatial filter atoms (responsible for spatial-only convolution) and their corresponding atom-coefficients (responsible for channel combination). Our observations reveal that solely fine-tuning ($\textit{Conform}$-ing) the spatial filter atoms, comprising of only a few hundred parameters, renders the transferability of the model efficient, without compromising on the predictive performance. Simultaneously, the static atom-coefficients serve to retain the base (foundational) knowledge from the pre-trained model. We rigorously assess this dual-faceted approach within the demanding and practical framework of cross-domain few-shot learning, showcasing the approach's substantial capability of transferring the knowledge in a parameter-efficient manner.

Primary Area: transfer learning, meta learning, and lifelong learning

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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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.

Submission Number: 7801