Rethinking Global Average Pooling: Your Classifier Is Secretly a Multi-Instance Learner

The paper presents a novel perspective on global average pooling (GAP) in image classifiers, proposing that these classifiers function as multi-instance learners (MIL) despite being trained with a single label per image. It demonstrates that the spatial class evidence remains accessible in multi-object scenes, allowing for the decomposition of image-level logits into a prediction grid, which aids in diagnosing classifier performance. This insight is significant for practitioners as it highlights potential improvements in model interpretability and robustness by addressing the limitations of mean aggregation in standard classification tasks.