Tweeted By @mblondel_ml

My paper Structured Prediction with Projection Oracles https://t.co/efAhlLEhoQ Key idea: using a Euclidean projection onto a convex set as output layer is guaranteed to make the squared loss equal or smaller than without projection. Comes with consistency/calibration guarantees. pic.twitter.com/dOqRXekuJT

— Mathieu Blondel (@mblondel_ml) December 5, 2019