Abstract
Progressively applying Gaussian noise transforms complex data distributionsto approximately Gaussian. Reversing this dynamic defines a generative model.When the forward noising process is given by a Stochastic Differential Equation(SDE), Song et al. (2021) demonstrate how the time inhomogeneous drift of theassociated reverse-time SDE may be estimated using score-matching. A limitationof this approach is that the forward-time SDE must be run for a sufficientlylong time for the final distribution to be approximately Gaussian. In contrast,solving the Schr\"odinger Bridge problem (SB), i.e. an entropy-regularizedoptimal transport problem on path spaces, yields diffusions which generatesamples from the data distribution in finite time. We present Diffusion SB(DSB), an original approximation of the Iterative Proportional Fitting (IPF)procedure to solve the SB problem, and provide theoretical analysis along withgenerative modeling experiments. The first DSB iteration recovers themethodology proposed by Song et al. (2021), with the flexibility of usingshorter time intervals, as subsequent DSB iterations reduce the discrepancybetween the final-time marginal of the forward (resp. backward) SDE withrespect to the prior (resp. data) distribution. Beyond generative modeling, DSBoffers a widely applicable computational optimal transport tool as thecontinuous state-space analogue of the popular Sinkhorn algorithm (Cuturi,2013).