The building blocks of a Big AI in healthcare | Sotirios A. Tsaftaris

Healthcare is under a perfect storm with AI/ML being offered as a solution to relieve at least one bottleneck: making decisions and predictions. Indeed, the detection of disease, segmentation of anatomy and other classical image analysis tasks, have seen incredible improvements due to deep learning. Yet these advances need lots of data: for every new task, new imaging scan, new hospital, new population, more training data are needed. The current paradigm of AI provides many tailored, laborious to train and develop, models. We need a new vision: a “Big AI” that can process several inputs, solve several tasks, generalize to new data, and make well-grounded predictions all relying on less supervision. I will advocate that central to paving the way for solutions that address such desiderata are better data representations learned without supervision. The presentation from a technical viewpoint will touch on multimodal learning, disentangled representation learning, meta-learning, semi- and weak-supervision, generative models, and causality.

https://www.dkfz.de/en/datascience/seminar/Tsaftaris.html