Comparing Probability Distributions with Conditional Transport

Explore the innovative concept of conditional transport (CT) as a new divergence for measuring differences between probability distributions in this seminar series talk. Delve into the introduction of amortized CT (ACT) and its applications in implicit distributions and stochastic gradient descent optimization. Learn how ACT utilizes two navigators to amortize the computation of conditional transport plans, providing unbiased sample gradients that are easy to calculate. Discover the benefits of applying ACT to generative model training, including its ability to balance mode covering and seeking behaviors while resisting mode collapse. Examine the improved performance achieved by substituting default statistical distances with ACT's transport cost in generative adversarial networks across various benchmark datasets. Gain insights from Associate Professor Mingyuan Zhou of the University of Texas at Austin, whose research spans machine learning, Bayesian statistics, and deep learning.