Weakly supervised learning

Last updated Apr 10, 2022 Edit Source

Weak supervision is a branch of machine learning where noisy, limited, or imprecise sources are used to provide supervision signal for labeling large amounts of training data in a supervised learning setting. This approach alleviates the burden of obtaining hand-labeled data sets, which can be costly or impractical. Instead, inexpensive weak labels are employed with the understanding that they are imperfect, but can nonetheless be used to create a strong predictive model.

See AI/Transfer learning, AI/Active learning and AI/Semi-supervised learning

• Other areas of machine learning exist that are likewise motivated by the demand for increased quantity and quality of labeled training data but employ different high-level techniques to approach this demand. These other approaches include AI/Active learning, AI/Semi-supervised learning, and AI/Transfer learning.
• Related to AI/One, few-shot learning. The most relevant problem to few-shot learning is weakly supervised learning with incomplete supervision where only a small amount of samples have supervised. By definition, weakly supervised learning with incomplete supervision includes only classification and regression, while few-shot learning also includes reinforcement learning problems. Moreover, weakly supervised learning with incomplete supervision mainly uses unlabeled data as additional information in E, while few-shot learning leverages various kinds of prior knowledge such as pretrained models, supervised data from other domains or modalities and does not restrict to using unlabeled data. Therefore, few-shot learning becomes weakly supervised learning problem only when prior knowledge is unlabeled data and the task is classification or regression.information.
• Weakly Supervised Learning: Introduction and Best Practices

• #PAPER A brief introduction to weakly supervised learning (2018)
• #PAPER A Graph-Based Method for Active Outlier Detection With Limited Expert Feedback (2019)

• In this case, only a (usually small) subset of training data is given with labels while the other data remain unlabeled (e.g., in image categorization the ground-truth labels are given by human annotators, and only a small subset of images can be annotated due to the human cost)
• #PAPER Learning from Incomplete and Inaccurate Supervision (Zhang 2021)

• In this case, only coarse-grained labels are given. Consider the image categorization task again. It is desirable to have every object in the images annotated; however, usually we only have image-level labels rather than object-level labels.
• #PAPER Labeled Data Generation with Inexact Supervision (Dai 2021)

• The given labels are not always ground-truth (e.g., the image annotator is careless, or some images are difficult to categorize)
• #PAPER Auxiliary Image Regularization for Deep CNNs with Noisy Labels (2016)
• #PAPER Anomaly detection with inexact labels (2019)