Efficient Algorithms for Generative Learning

Obiettivo

"Scientific progress in the field of computer vision has traditionally developed along two largely separate paths. One focuses on image analysis tasks such as classification and reconstruction, the other on image synthesis tasks such as inpainting and generation. This separation reveals a fundamental challenge in this field, which we refer to as the analysis–synthesis gap. A major reason for this artificial separation is the strong focus of the scientific community on achieving state-of-the-art, task-specific performance on public benchmarks and applications. While this has led to remarkable progress, it has also hindered the development of unified approaches that seamlessly combine analysis and synthesis.

Motivated by ""analysis-by-synthesis"", a core principle of cognitive science and statistical pattern theory, we argue that any probabilistic vision model is robust and falsifiable only if it possesses the dual ability to analyze and generate images. Building on this insight, our proposed research project EAGLE aims to close the gap between analysis and synthesis by developing efficient generative models and sampling algorithms within a Bayesian framework that naturally integrates these two interconnected capabilities.

The outcomes of the project are twofold: (i) fully Bayesian models grounded in state-of-the-art neural architectures; and (ii) a new family of accelerated sampling algorithms tailored to high-dimensional problems in computer vision. These innovations combine theoretical rigor with practical computational efficiency, defining a new state of the art for solving inverse problems in computer vision, with applications ranging from image denoising to medical image reconstruction. Beyond vision, the project’s outcomes extend to other scientific fields, such as protein engineering, where Bayesian formulations, generative modeling, and sampling play equally critical roles."