“System-Embedded Diffusion Bridge Models”
Approximating mappings between clean and distorted data samples enables the solution of various complex inverse problems frequently encountered in real-world scenarios. To this end, recent ground-breaking research has leveraged the framework of score-based modeling in two orthogonal ways: an unsupervised approach using pretrained diffusion models and a supervised one using so-called diffusion bridges. In this talk, I will present an approach that combines the strengths of both by employing a relatively overlooked matrix-valued formulation of score-based modeling based on stochastic differential equations, which embeds the linear measurement model directly into its coefficients.
Bio
I am a first-year PhD student under the supervision of prof. dr hab. inż. Przemysław Biecek at the University of Warsaw, conducting research at the intersection of interpretability and generative modeling. I am passionate about mathematics and its applications in shaping the world around us. I believe that, just as with physics in the last century, fully understanding the inner workings of AI requires harnessing the power of the advanced mathematical tools developed over centuries of research.