Enhancing out-of-distribution detection via diversified multi-prototype contrastive learning

Yulong Jia, Jiaming Li, Ganlong Zhao, Shuangyin Liu, Weijun Sun, Liang Lin, Guanbin Li

Abstract: Highlights•A multi-prototype contrastive learning method preserves semantic structures within each class.•An activation constraint module reduces the impact of extreme activations during inference.•Extensive experiments on OOD benchmarks confirm the superiority of our method over state-of-the-art.