Reshaping Model Output Space Via Deep Kernel Density Estimation Networks

Lior Katz, Tomer Melamed

Keywords: Deep KDE, probability density transformations, Kernel Density Estimation

Abstract:

Traditional classification models are typically optimized solely for their specific training task without considering the properties of the underlying probability distribution of their output space. As the use of these models for downstream tasks becomes more prevalent, it becomes advantageous to have a framework that can transform the output space of such models to a more convenient space without
sacrificing performance. In this paper, we introduce DeepKDE, a novel method which enables the transformation of arbitrary output spaces to match more desirable distributions, such as Normal and Gaussian Mixture Models. We explore the properties of the new method and test its effectiveness on ResNet-18 and vision transformers trained on CIFAR-10 and Fashion MNIST datasets. We show that DeepKDE models succeed in transforming the output spaces of the original models while outperforming them in terms of accuracy.

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.

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: 7940