Are you Exploiting Your Assumptions? Towards Effective Priors for Biomarker Discovery and Functional Predictions

Are you Exploiting Your Assumptions? Towards Effective Priors for Biomarker Discovery and Functional Predictions

Bayesian modeling is a powerful framework to handle stochasticity present in the world via probability theory; its benefits include parameter sharing, model averaging (via integration), and uncertainty estimation. Yet, encoding human assumptions into priors is far from trivial. In this talk, I will illustrate the potential benefits of prior design in two specific scenarios. In the first half of the talk, I will present a Bayesian non-parametric latent feature model for heterogeneous datasets. The proposed model is useful to identify meaningful biomarkers in clinical trials via latent variables; I will show results on a real application for an immunotherapy treatment for liver cancer (https://doi.org/10.1186/s12885-019-5472-0). In the second half of the talk, I will introduce a novel prior for Bayesian neural networks that allows for direct specification of functional desiderata in terms of amplitude variance and lengthscale. We show consistency of the estimated regression function, as well as functional properties leading to better generalization.

[Slides]

日期：
演讲者：
Melanie Fernandez Pradier
所属机构：
Harvard University