Abstract:
|
Deep learning, neural networks (NN), random forests (RF) and Bayesian Additive Regression Trees (BART) are popular in industry and receive much attention in academia as flexible and highly scalable machine learning models. As powerful as methods such as NN are, they largely remain constrained to the task of point prediction. When applications call for probabilistic information, statistical models such as BART reign supreme. However, the ability of statistical models to scale to the size and complexity of datasets currently usable in machine learning remains a challenge. Recent developments in Bayesian regression tree models aim to relax these constraints.
This roundtable is an opportunity for academic and industry stakeholders to discuss (a) what an open-source software ecosystem for tree-based models should look like, what key methodologies should be made available, and what scalable computing infrastructure should be targeted so that a high-quality reference implementation of these modern Bayesian regression methods can be made available to the community; and (b) what research directions and developments are most needed.
|