3IA PhD/Postdoc Seminar #49

Flash presentation - When are Two Scores Better than One? Investigating Ensemble of Diffusion Models

Abstract: Diffusion models now generate high-quality, diverse samples, with an increasing focus on more powerful models. Although ensembling is a well-known way to improve supervised models, its application to unconditional score-based diffusion models remains largely unexplored. In this work we investigate whether it provides tangible benefits for generative modelling. We find that while ensemble generally improves the score-matching loss and model likelihood, it fails to consistently enhance perceptual quality metrics such as FID. Our study spans across a breadth of aggregation rules using Deep Ensembles, Monte Carlo Dropout, and Random Forests on CIFAR-10, FFHQ, and tabular data. We attempt to explain this discrepancy by investigating possible explanations, such as the link between score
estimation and image quality. Finally, we provide theoretical insights into the summing of score models, which shed light not only on ensembling but also on several model composition techniques (e.g. guidance).

Flash presentation - ShapeShifter: 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion

Abstract: We present ShapeShifter, a new 3D generative model that learns to synthesize shape variations based on a single example. While generative methods for 3D objects have recently attracted much attention, current techniques often lack geometric details and/or require long training times and large resources. Our approach remedies these issues by combining sparse voxel grids and multiscale point, normal, and color sampling within an encoder-free neural architecture that can be trained efficiently and in parallel. We show that our resulting variations better capture the fine details of their original input and can capture more general types of surfaces than previous SDF-based methods. Moreover, we offer interactive generation of 3D shape variants, allowing more human control in the design loop if needed.

Constructions in mathematics with a little help from AI

Abstract: In this seminar, I will introduce PatternBoost, a novel approach by Charton, Ellenberg, Wagner and Williamson (2024), that combines traditional search algorithms with neural networks to discover interesting mathematical constructions (arXiv:2411.00566). I will begin with an overview of transformers, explaining how they work. Then, I will delve into how PatternBoost leverages transformers in its two-phase approach: a local phase, where classical search algorithms generate promising constructions, and a global phase, where a transformer is trained on the best constructions to generate new candidates. Finally, I will present concrete examples where PatternBoost has successfully solved problems in extremal combinatorics, particularly in graph theory.

Open discussion about all the contributions