Deep Learning, Prediction, and Validation: Innovations in Statistical Modeling and Applications to Medical/Health Big Data (ADDED FEE) — Professional Development Continuing Education Course
ASA, Biopharmaceutical Section
We begin with a review of (a) recent advances in computer vision and deep learning and how it links with statistical/machine learning, (b) the underlying statistical theories of convolutional neural networks, gradient descent, graphical models, and hidden Markov random fields, and (c) AI applications to medical imaging and automated analysis of electronic medical and health data. Whereas high-performance computing and advanced programming have overcome the computational hurdles in the analysis of "big data" for prediction and classification, we next describe how statistical innovations have provided major breakthroughs in the validation of scientific theories based on complex experimental data in biomedical and astrophysics. Because big data typically require variable/hypothesis selection based on some sparsity assumption to make the inference problem feasible, there is contemporaneous awareness of irreproducible research in modern science. In particular, novel statistical methods in post-selection inference and hybrid resampling are presented to address this "reproducible (replication) crisis".