Open-World Semi-Supervised Learning

ICLR 2022  ·  Kaidi Cao, Maria Brbic, Jure Leskovec ·

A fundamental limitation of applying semi-supervised learning in real-world settings is the assumption that unlabeled test data contains only classes previously encountered in the labeled training data. However, this assumption rarely holds for data in-the-wild, where instances belonging to novel classes may appear at testing time. Here, we introduce a novel open-world semi-supervised learning setting that formalizes the notion that novel classes may appear in the unlabeled test data. In this novel setting, the goal is to solve the class distribution mismatch between labeled and unlabeled data, where at the test time every input instance either needs to be classified into one of the existing classes or a new unseen class needs to be initialized. To tackle this challenging problem, we propose ORCA, an end-to-end deep learning approach that introduces uncertainty adaptive margin mechanism to circumvent the bias towards seen classes caused by learning discriminative features for seen classes faster than for the novel classes. In this way, ORCA reduces the gap between intra-class variance of seen with respect to novel classes. Experiments on image classification datasets and a single-cell annotation dataset demonstrate that ORCA consistently outperforms alternative baselines, achieving 25% improvement on seen and 96% improvement on novel classes of the ImageNet dataset.

PDF Abstract ICLR 2022 PDF ICLR 2022 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Open-World Semi-Supervised Learning CIFAR-10 ORCA (ResNet-18) Seen accuracy (50% Labeled) 88.2 # 4
Novel accuracy (50% Labeled) 90.4 # 3
All accuracy (50% Labeled) 89.7 # 4
Seen accuracy (10% Labeled) 82.8 # 3
Novel accuracy (10% Labeled) 85.5 # 3
All accuracy (10% Labeled) 84.1 # 3
Open-World Semi-Supervised Learning CIFAR-100 ORCA (ResNet-18) Seen accuracy (10% Labeled) 52.5 # 3
Novel accuracy (10% Labeled) 31.8 # 3
All accuracy (10% Labeled) 38.6 # 3
Seen accuracy (50% Labeled) 66.9 # 1
Novel accuracy (50% Labeled) 43.0 # 1
All accuracy (50% Labeled) 48.1 # 1
Open-World Semi-Supervised Learning ImageNet-100 ORCA (ResNet-50) Seen accuracy (50% Labeled) 89.1 # 1
Novel accuracy (50% Labeled) 72.1 # 2
All accuracy (50% Labeled) 77.8 # 3
Seen accuracy (10% Labeled) 83.9 # 1
Novel accuracy (10% Labeled) 60.5 # 2
All accuracy (10% Labeled) 69.7 # 1
Novel Object Detection LVIS v1.0 val ORCA Cao et al. (2022) Novel mAP 0.49 # 4
Known mAP 20.57 # 4
All mAP 2.03 # 4

Methods


No methods listed for this paper. Add relevant methods here