Hide&Seek: Learning to Explain in an End-to-End Differentiable Network
Hide&Seek replaces hard feature removal with a differentiable substitute so selection and prediction can train together.
The paper targets instance-wise feature selection, where the important inputs can change from example to example. Its authors say prior selector-predictor methods can suffer from information leakage and slower training because parts of the process are not fully differentiable. Hide&Seek uses a single objective for feature selection and prediction, replacing proportions of features instead of discretely dropping them. The authors report state-of-the-art results across experiments and faster training, with stability helped by parsimony-weight annealing. ArXiv · AI/CL/LG's note
The paper targets instance-wise feature selection, where the important inputs can change from example to example. Its authors say prior selector-predictor methods can suffer from information leakage and slower training because parts of the process are not fully differentiable. Hide&Seek uses a single objective for feature selection and prediction, replacing proportions of features instead of discretely dropping them. The authors report state-of-the-art results across experiments and faster training, with stability helped by parsimony-weight annealing. ArXiv · AI/CL/LG's note
score 4