Abstract
Recent advance in contrastive learning has shown remarkable performance.However, the vast majority of approaches are limited to the closed-worldsetting. In this paper, we enrich the landscape of representation learning bytapping into an open-world setting, where unlabeled samples from novel classescan naturally emerge in the wild. To bridge the gap, we introduce a newlearning framework, open-world contrastive learning (OpenCon). OpenCon tacklesthe challenges of learning compact representations for both known and novelclasses, and facilitates novelty discovery along the way. We demonstrate theeffectiveness of OpenCon on challenging benchmark datasets and establishcompetitive performance. On the ImageNet dataset, OpenCon significantlyoutperforms the current best method by 11.9% and 7.4% on novel and overallclassification accuracy, respectively. We hope that our work will open up newdoors for future work to tackle this important problem.