Enhanced Multi-Instance Partial Label Learning via Average Gradient Outer Product

Multi-instance partial-label learning (MIPL) is a recently proposed dual weakly supervised learning framework where each training bag is annotated with a candidate label set containing one true label and several false positives. The key challange of MIPL problem is that the relations between the key instances and the ground-truth labels are much more covered due to the false positive labels. Existing methods usually rely on the model's own predictions and roughly aggregate instance features according to the learned attention weights, which is easily misled by false positives and lets key instances be overlooked. Here, we propose Average Gradient Outer Product based Multi-instance Partial-Label Learning~(AGOPMIPL) method, where the average gradient outer product (AGOP) is directly calculated from the bag-level features and the model prediction, then it is integrated in an attention module to amplify discriminative feature directions and thereby helps key-instance identification. Moreover, the feature prototypes and a progressive disambiguation strategy are introduced to further suppress noisy candidates. The experimental studies on four MIPL benchmarks and the real-world \textsf{CRC-MIPL} dataset are performed and AGOPMIPL consistently outperforms five state-of-the-art baselines, with up to $25.9\%$ relative gain on \textsf{CRC-MIPL-KMeansSeg}.

Enhanced Multi-Instance Partial Label Learning via Average Gradient Outer Product (ICML 2026, submission 25270) Machine-learning systems usually need large amounts of carefully labeled data, but accurate labels are expensive. In many real settings the labels are doubly imperfect. Take an image labeled by several crowd workers: the image is really a collection of small regions, only some of which matter, and the workers may disagree, so the image arrives with a short list of candidate labels of which only one is correct. A method has to figure out at the same time which regions are important and which candidate label is the true one, and the wrong candidates can easily mislead it. We introduce AGOPMIPL, which handles both problems with a single idea. Rather than trusting the noisy labels directly, it looks across the entire training set to find which characteristics the model consistently relies on when making decisions, and then reshapes the data to emphasize exactly those characteristics. This steers the model's attention onto the regions that genuinely determine the label, and it gradually sorts the true label out from the candidates as training proceeds. On standard benchmarks and a real colorectal-cancer image dataset labeled by non-experts, AGOPMIPL is consistently more accurate than the best existing methods, improving accuracy by up to about 26% in the hardest real-world case. This could make learning from cheap, crowd-sourced, or otherwise weakly labeled data more reliable.