Review for NeurIPS paper: Learning to Select Best Forecast Tasks for Clinical Outcome Prediction

Summary and Contributions: The authors propose a method that learns a model that can classify a given multi-variate time series. Their learned system is a recurrent neural network (RNN) that maps a multi-variate time series (which represents an individual patient) into a vector, and a classifier that maps that vector to one of C possible classes. They propose an algorithm to learn the weights of the RNN, the classifier, and the weights. They validate their proposed algorithm in a clinical dataset, where each patient is defined by a 96-dimensional time series. Each of the 96 time series represents the measurement of a variable over time (e.g.