Structure and Behavior in Weight Space Representation Learning

Léo Meynent, Ivan Melev, Konstantin Schürholt, Goeran Kauermann, Damian Borth

Keywords: weight space learning, hyper-representations, deep weight spaces, representation learning, model reconstruction, model weights generation

TL;DR: We demonstrate how the combination of structural and behavioral losses improve the training of hyper-representation autoencoders in weight space.

Abstract: The weights of neural networks (NNs) have recently gained prominence as a new data modality in machine learning, with applications ranging from accuracy and hyperparameter prediction to representation learning or weight generation. One approach to leverage NN weights involves training autoencoders (AEs) with contrastive and reconstruction losses. Indeed, such models can be applied to a wide variety of downstream tasks, and they demonstrate strong predictive performance and low reconstruction error. However, despite the low reconstruction error, these AEs reconconstruct NN models that fail to match the performance of the original ones. In this paper, we identify a limitation of weight-space AEs, specifically highlighting that structural weight reconstruction alone fails to capture some features critical for reconstructing high-performing models. To address this issue, we propose a behavioral loss for training AEs in weight space. This behavioral loss focuses on the features essential for reconstructing performant models, which are not adequately captured by structural reconstruction. We evaluate the capabilities of AE trained using this novel loss on three different model zoos: we demonstrate that when combining structural and behavioral losses, we can reconstruct and generate models that match the performance of the original models. With our exploration of representation learning in deep weight spaces, we show that a strong synergy exists between structural and behavioral features, and that combining them results in increased performance across all evaluated downstream tasks.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation 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.

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: 10732