Score-Based Generative Modeling with Hamilton-Jacobi-Bellman Regularizers
- Presenter
- Benjamin Burns
- Campus
- UMass Amherst
- Sponsor
- Markos Katsoulakis, Department of Mathematics and Statistics, UMass Amherst
- Schedule
- Session 2, 11:30 AM - 12:15 PM [Schedule by Time][Poster Grid for Time/Location]
- Location
- Poster Board A41, Campus Center Auditorium, Row 3 (A41-A60) [Poster Location Map]
- Abstract
We introduce a class of mathematically justified regularizers for training score-based generative models (SGM) based on its connections to mean-field games and Hamilton-Jacobi-Bellman (HJB) equations. Recently it has been shown that score-based generative models can be formulated in terms of a mean-field game (MFG). This MFG formulation admits alternate characterizations of the SGM based on its optimality conditions, which are a set of coupled nonlinear partial differential equations (PDE). The first PDE is a controlled Fokker-Planck equation which is equivalent to the denoising process in SGM, while the second PDE is a HJB equation that characterizes the optimal controller, whose solution is related to the score function. Based on this mathematically rigorous connection, we propose a new objective function for the score function, and develop a PDE-based theory for explaining and understanding the role of latent space in SGM. In addition to the implicit score-matching objective, we include two regularizers, the first based on measuring the discrepancy in the HJB equation, the second based on measuring discrepancy in satisfying the terminal condition. The terminal condition is determined by the training data. By including these regularizers, we are informing the structure of the score function, thereby constraining the search space. This strategy is well-grounded through the theory of mean-field games and HJB equations. Through experiments we will show that the HJB regularizer helps learn the score function in a more stable way, while the terminal condition regularizer is associated with higher quality samples.
- Keywords
- machine learning, generative modeling, mean-field games, Hamilton-Jacobi-Bellman equation
- Research Area
- Mathematics and Statistics
SIMILAR ABSTRACTS (BY KEYWORD)
Research Area |
Presenter |
Title |
Keywords |
Probability, Statistics, and Machine Learning |
Waghe, Shreyas |
|
Fair Machine Learning (0.758621), Machine Learning (0.846154)
|
Engineering |
Liousas, Demetri |
|
machine learning
|
Computer Science |
Thornton, Peter David |
|
machine learning
|
Computer Science |
Berduo, Alan Jesse |
|
Machine learning
|
Engineering |
Li, Agnes |
|
Machine learning
|