How do you choose the appropriate level and type of supervision for weakly supervised learning tasks?

2 Label quantity

Another factor to consider when choosing the supervision level for WSL is the quantity of the labels. Label quantity refers to the proportion of the data that has labels, as well as the number of labels per data point. For example, some WSL tasks may involve only a small fraction of labeled data, while others may have multiple labels for each data point. The label quantity can affect the complexity and feasibility of the WSL models, as well as the amount of supervision required. Generally, the higher the label quantity, the less supervision is needed to learn from the data. However, too many labels can also be redundant or conflicting, especially if they come from different sources or have different levels of quality. Therefore, it is important to evaluate the label quantity and choose a supervision level that maximizes the information and diversity of the labels.

3 Label type

A third aspect to consider when choosing the supervision level for WSL is the type of the labels. Label type refers to the format and granularity of the labels, as well as the relation between the labels and the data. For example, some WSL tasks may involve binary or categorical labels, while others may have numeric or ordinal labels. Some WSL tasks may also have point-wise or instance-level labels, while others may have group-wise or set-level labels. Some WSL tasks may also have direct or explicit labels, while others may have indirect or implicit labels. The label type can affect the expressiveness and flexibility of the WSL models, as well as the amount of supervision required. Generally, the more complex and diverse the label type, the more supervision is needed to learn from the data. However, too much supervision can also limit the generality or adaptability of the WSL models, especially if the labels are not aligned with the target domain or task. Therefore, it is important to determine the label type and choose a supervision level that matches the learning objectives and constraints.

4 Supervision methods

The final aspect to consider when choosing the supervision level for WSL is the methods or techniques that can be used to provide or enhance the supervision. Supervision methods can be classified into three main categories: data-level, model-level, and task-level methods. Data-level methods aim to improve the quality or quantity of the labels by using techniques such as data augmentation, data cleaning, data fusion, or data generation. Model-level methods aim to improve the robustness or flexibility of the models by using techniques such as regularization, multi-task learning, semi-supervised learning, or self-training. Task-level methods aim to improve the effectiveness or efficiency of the tasks by using techniques such as active learning, curriculum learning, transfer learning, or meta-learning. The choice of supervision methods can depend on the characteristics and requirements of the data, the models, and the tasks, as well as the availability and feasibility of the techniques. Therefore, it is important to explore and compare different supervision methods and choose a supervision level that optimizes the performance and resource utilization of the WSL.