InfoDPP-PAC: Principled Patch Selection for Whole Slide Image Analysis

What the math gives to ML

The paper combines Gaussian-process posterior uncertainty with determinantal diversity, turning subset selection into an information-gain problem rather than a classifier-specific attention heuristic. The transferable asset is a decomposable objective whose log-determinant term rewards genuinely new information and whose greedy marginal gains naturally support variable-length selection. A practical neural-network adaptation is to use a frozen teacher to score candidate tokens, patches, or retrieved items, fit a GP covariance in embedding space, and greedily retain items while the posterior information gain remains above a calibrated threshold. This can reduce attention or memory costs while preserving coverage, although the GP and DPP machinery itself is established and the main novelty is the principled integration and stopping rule.