Central Force Field: Unifying Generative and Discriminative Models While Harmonizing Energy-Based and Score-Based Models

Guangrun Wang, Chen Lin, Philip Torr

Keywords: Central Force Field, Generative Model, Discriminative Model, Energy-Based Model, Score-Based Model

Abstract:

In the pursuit of Artificial General Intelligence, a prevalent approach is to establish a comprehensive unified foundation model that addresses multiple tasks concurrently. However, creating such a model that unifies generative and discriminative models presents significant challenges.
This paper aims to realize this unified model aspiration by suggesting the incorporation of a central force field from physics.
More precisely, within the framework of this central force field, the potential functions governing the data distribution and the joint data-label distribution become intricately interwoven with a standard discriminative classifier, rendering them well-suited for handling discriminative tasks.
Moreover, the central force field exhibits a captivating characteristic: objects located within this field experience an attractive force that propels them towards the center. This phenomenon of centripetal motion, orchestrated by the force field, has the remarkable capability to progressively revert diffused data to its original configuration, thereby facilitating the execution of generative tasks.
Our proposed method adeptly bridges the realms of energy-based and score-based models. Extensive experimental validation attests to the effectiveness of our approach, showcasing not only its prowess in image generation benchmarks but also its promising competitiveness in image classification benchmarks.

Primary Area: general machine learning (i.e., none of the above)

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/2024/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: 5509