Event Date: October 29, 2020 16:15
Semi-Structured Deep Distributional Regression
SFB876 & DoDSc guest
Abstract:
Semi-Structured Deep Distributional Regression (SDDR) is a unified network architecture for deep distributional regression in which entire distributions can be learned in a general framework of interpretable regression models and deep neural networks. The approach combines advanced statistical models and deep neural networks within a unifying network, contrasting previous approaches that embed the neural network part as a predictor in an additive regression model. To avoid identifiability issues between different model parts, an orthogonalization cell projects the deep neural network part into the orthogonal complement of the statistical model predictor, facilitating both estimation and interpretability in high-dimensional settings. The framework is implemented in an R software package based on TensorFlow and provides a formula user interface to specify the models based on the linear predictors. In the second part of the talk, models in which tasks are represented as direct acyclic graphs (DAGs) are considered, and methods for guaranteeing both timing constraints and memory feasibility are presented. In particular, solutions for bounding the worst-case memory space requirement for parallel tasks running on multi-core platforms with scratchpad memories are discussed.
Short bio:
Dr. David Rügamer is a lecturer and postdoctoral research fellow at the chair of Statistical Learning and Data Science (Prof. Bischl), Department of Statistics, LMU Munich, where he also leads two research subgroups on machine learning and deep learning. Before joining the chair, he has worked as Senior Data Science in the industry with focus on data engineering and deep learning research. From 2014 to 2018 he did his PhD under the supervision of Prof. Dr. Sonja Greven and was partly funded by the Emmy Noether project ‘Statistical Methods for Longitudinal Functional Data’.
