Work with @zhezeng0908 @KareemYousrii @guyvdb ! arXiv link: and github link:

We evaluate our approach on three well-known settings: Learning Label Proportions (LLP), Multiple Instance Learning (MIL), and Positive Unlabeled Learning (PU) over a wide variety of datasets, showing SOTA performance 👏👏👏.

Our algorithm reasons over the entire count distribution, enabling range queries over counts of the form at least k or at most k 🤯. We can then utilize the answer to these probabilistic queries to optimize task-dependent objective functions and train models end to end 😎.

Take Learning Label Proportions, where given count of positive instances in a bag, we want to optimize the bag posterior. How do we do this 🤔? Enumerating all labelings is intractable 👎 and approximates are probabilistically unsound 👎. Our answer: dynamic programming 🦸!!!.

Many approaches to weakly supervised learning are ad hoc, inexact, and limited in scope 😞. We propose Count Loss 🎉, a simple ✅, exact ✅, differentiable ✅, and tractable ✅ means of unifying count-based weakly supervised settings! See at NeurIPS 2023!.