Review for NeurIPS paper: Learning from Label Proportions: A Mutual Contamination Framework

The work is strongly based on the results for mutual contamination models, which is specially designed and discussed for binary classification problems. As a result, the benchmark approaches involved in the experimental section are confined in two early proposed methods for binary problems. Many up-to-date models that focus on multi-class LLP problems are not touched by this work. In other words, there is a gap between this work and multi-class LLP problem. In particular, thanks to deep neural networks, the performance on LLP has been greatly improved by recent work, such as the work in [1], [11], [18], [31], and [34], on much complicated datasets, e.g., image data.