Title : Bayesian Deep Learning with Gaussian Processes
Venue: IEOR Seminar Hall (2nd floor)
Date: 14th November 2018
Time: 3.00 pm to 4.30 pm
Abstract: Bayesian learning allow the incorporation of prior domain knowledge in probabilistic models and provides uncertainty estimates which are useful for decision making. Bayesian non-parametrc models additionally allow one to learn rich and flexible models by letting the model complexity to grow with the data. They help to overcome the problem of model selection to a great extent. In this talk, I will introduce Bayesian non-parametric model, Gaussian processes, which provides a flexible framework to model functions. I will discuss how deep Gaussian proceses improves the function learning capacity and overcome many limitations of popular deep learning models. Further, I will also present our recent work on convolutional deep Gaussian processes for image classification problems and the use of natural gradients to improve the convergence in deep Gaussian process training.
Speaker Profile: https://sites.google.com/site/pksrijith/