Robust learning from untrusted sources
Robust learning from untrusted sources Konstantinov & Lampert, ICML’19
Welcome back to a new term of The Morning Paper! Just before the break we were looking at selected papers from ICML’19, including “Data Shapley.” I’m going to pick things up pretty much where we left off with a few more ICML papers…
Data Shapley provides us with one way of finding and correcting or eliminating low-value (and potentially harmful) data points from a training set. In today’s paper choice, Konstantinov & Lampert provide a way of assessing the value of datasets as a whole. The idea is that you might be learning e.g. a classifier by combining data from multiple sources. By assigning a value (weighting) to each of those data sources we can intelligently combine them to get the best possible performance out of the resulting trained model. So if you need more data in order to improve the performance of your model, ‘Robust learning from untrusted sources’ provides an approach that lets you tap into additional, potentially noisier, sources.
It’s similar in spirit to Snorkel which we looked at last year, and is designed to let you incorporate data from multiple ‘weakly supervised’ (i.e. noisy) Continue reading
