Memory Banks and Sharper Angles Give AI a Keener Eye for Look-Alike Images

The second component is a first-in-first-out memory bank with capacity K times the mini-batch size.

The experimental evidence spans five primary benchmarks, CIFAR-10, CIFAR-100, Tiny-ImageNet, CUB200-2011, and Stanford Dogs, plus a supplementary small-data setting built from PASCAL VOC 2005, evaluated with ResNet-18, ResNet-50, ConvNeXt-T, and two TinyViT transformers.

The authors also replaced the two-view augmentation strategy used by most contrastive methods with single-view encoding, letting the memory bank supply the extra comparison samples.

Future work will explore adaptive or learnable kappa and more sophisticated memory mechanisms such as staleness-aware weighting.

Sigmoid-based supervised contrastive learning with von mises–fisher similarity and memory banks for visual recognition.