About event:
We are pleased to invite you to an upcoming research talk on:“Some Mathematical Aspects of Score-Based Generative Models”
Diffusion probabilistic models have emerged as state-of-the-art tools in generative modeling, capable of producing high-resolution samples from complex, high-dimensional distributions such as images.
Despite their effectiveness, these models have certain drawbacks:● Unlike variational autoencoders, they maintain high dimensionality throughout the generation process.● They are susceptible to memorizing training data.
This talk focuses on the second issue by exploring score-based generative modeling.
We introduce a smoothed empirical score and derive improved bounds on the KL-divergence between the true data distribution and the approximation generated via this score.
Our proposed estimator reduces overfitting and enhances generalization.
These results will be supported by experiments on both synthetic and real datasets.
Speaker:
Maria Han Veiga - Department of Mathematics, Ohio State University
Language:
English