Abstract: Machine learning models deployed in the wild naturally encounter unlabeled samples from both known and novel classes. Challenges arise in learning from both the labeled and unlabeled data, in an open-world semi-supervised manner. In this paper, we introduce a new learning framework, open-world contrastive learning (OpenCon). OpenCon tackles the challenges of learning compact representations for both known and novel classes and facilitates novelty discovery along the way. We demonstrate the effectiveness of OpenCon on challenging benchmark datasets and establish competitive performance. On the ImageNet dataset, OpenCon significantly outperforms the current best method by 11.9% and 7.4% on novel and overall classification accuracy, respectively. Theoretically, OpenCon can be rigorously interpreted from an EM algorithm perspective—minimizing our contrastive loss partially maximizes the likelihood by clustering similar samples in the embedding space. The code is available at https://github.com/deeplearning-wisc/opencon.
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: We have addressed the reviewers’ comments and concerns in individual responses to each reviewer. The reviews allowed us to strengthen our manuscript in the camera-ready version and the major changes made are summarized below: + [R1] Added results on fine-grained datasets in Appendix N. + [R1] Revised framing of representation learning in abstract and introduction. + [R1] Added baselines of k-means, Rankstat+, and UNO+ in Table 3. + [R1, R3] Revised Section 1, Section 2, and Table 1 in blue, highlighting GCD and ORCA. + [R2, R4] Added the ImageNet-1k Experiments in Appendix L. + [R2] Added Labeling Ratio = 0.5, $y_l$= 50 results to Table 6. + [R2] Updated Figure 3 with ablations on both CIFAR-100 and ImageNet-100. + [R3] Added the discussion on weakness in Section 8, and revised broader impact in Appendix A. + [R3] Changed the title to "OpenCon: Open-world Contrastive Learning". + [R3] Added a remark on the summary of Appendix E in Section 3.2 + [R4] Added the core contribution discussion w.r.t. PCL in Appendix M. (\* As abbreviations, we refer to **Reviewer MY9C** as R1, **Reviewer wCPe** as R2, **Reviewer 6dAv** as R3, and **Reviewer 6PFD** as R4 respectively.)
Assigned Action Editor: ~Neil_Houlsby1
Submission Number: 478