Computer Vision News - March 2020

3 Summary Klaus Maier-Hein 11 with supervised learning to learn from partly annotated data. However, most of the approaches that are currently being developed involve learning more or less from scratch. “What we are not very good at yet is building up knowledge from different sources and distilling it,” he explains . “It’s a thing that humans are able to do. Some humans have a knowledge representation that is really adaptive and that helps them understand new problems very fast and learn very quickly. With machine learning, we are often stuck with our approach of initializing our method with random parameters and then training millions of parameters from scratch for each problem.” Whi le it is certainly not easy to transfer knowledge from one problem to another, even the transfer of a principle architecture or method is not always straightforward. Especial ly in medicine, there’s a lot of heterogeneity in the datasets and so this can lead to a lot of noise. Currently, he says, you might even have to hire a PhD student and invest a couple of years’ time in order to make the transition. But surely this represents exciting times for young researchers looking for emerging f ields to enter and make cutting-edge progress? Yes and no, Klaus responds. The sheer abundance of papers out there – about 100 per day on deep learning being publ ished on arXiv - make it very hard to f ind the signal in the noise. “What we are not very good at yet is building up knowledge from different sources and distilling it” Tier 1 Shared Data Tier 3 Federated Data ✔ Large-scale joint repository of high-quality annotated data Tier 2 Synthesized Data ✔ Major obstacles resolved by innovative federated infrastructure Shared Data public non- public Sharable Data Dark Data ▪ Majority of data not available for AI ▪ Technical, organisational & ethical obstacles Medical image computing at scale ✔ Large-scale joint repository of high-quality annotated data ✔ Major obstacles resolved by innovative federated infrastructure