Uncertainty Estimation and Generalization Bounds for Modern Deep Learning

In three linesThesis on Bayesian principles applied to deep neural networks. Introduces DVIP (Deep Variational Implicit Process) for scalable Bayesian inference, and two post-hoc methods (VaLLA, FMGP) to calibrate uncertainty on pretrained networks. Develops unified theoretical framework connecting diversity, smoothness, and stochasticity via PAC-Bayesian and large-deviation theory.

Summary generated by Claude — human-verified