Discrimination for Generation

Junhoo Lee, Hyeonjin Kim, Han Sangbum, Nojun Kwak

Keywords: Generative model, Discriminative model, Neural Tangent Kernel, Functional Analysis

TL;DR: This paper first presents a general schema to convert discriminative models to generative models.

Abstract:

There are two primary approaches to learning from data: discriminative models, which make predictions based on provided data, and generative models, which learn data distributions to create new instances. This paper introduces a novel framework, Discrimination for Generation (DFG), as the first attempt to bridge the gap between discriminative and generative models. Through DFG, discriminative models can function as generative models. We leverage the Neural Tangent Kernel (NTK) to map discriminative models into a connected functional space, enabling the calculation of the distance between the data manifold and a sampled data point.
Our experimental results demonstrate that the proposed algorithm can generate high-fidelity images and can be applied to various tasks such as Targeted Editing and Inpainting, in addition to both unconditional and conditional image generation.
This connection provides a novel perspective for interpreting models. Moreover, our method is algorithm-, architecture-, and dataset-agnostic, offering flexibility and proving to be a robust technique across a wide range of scenarios.

Primary Area: generative models

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: 9521