Weak Supervision is a machine learning method used in supervised learning using noisy, limited, and imprecise data sources for labelling training data. Weak supervision is useful when there are large amount of low-quality labels available and relevant to a machine learning task which can help make higher levels of abstraction more efficient in creating strong predictive machine learning models. It can also help machine learning engineers gain a better understanding of unlabeled data such as heuristics, distant supervision, constraints, expected distributions, and invariances.
Weak supervision can be used when there is only a few strong labels available for training machine learning systems which are targeting the actual desired prediction when there are many weak labels available in the training data. The weak labels do not directly target the prediction being made but are still useful to developing more accurate prediction outcomes. This makes weak supervision useful for accelerating the learning of strong tasks using strong labels by including training on the weak labels as part of the machine learning process. The amount of acceleration is dependent on the amount of available weak labels, and on their relation to the strong labels and the prediction being trained.
Incomplete weak supervision deals with datasets containing small amounts of labeled data (which is not sufficient for training models) and the availability of abundant unlabelled training data. There are two techniques for making sense of incomplete weak supervision data called active learning and semi-supervised learning.
Active learning involves a human expert that helps find important ground-truth labels for unlabelled data, and semi-supervised learning attempts to exploit data within labelled and unlabelled data in a way which improves learning and training performance without human intervention.
There are two types of semi-supervised learning which take into account different assumptions about training data. There is transductive learning that assumes a 'closed-world' model treating all unlabelled data as test data for optimizing training performance, and pure semi-supervised learning that assumes an 'open-world' model assuming that some test data is unknown and all unlabelled data should not necessarily be treated as test data for training purposes.
Inexact weak supervision is based on training models with some supervision information present, but not the exact amount desired. This form of weak supervision is used for helping to solve problems such as drug discovery where researchers can only provide supervised learning using information related to known molecules.
Inexact weak supervision takes advantage of multi-instance learning; a type of supervised machine learning based on aggregating instances in relation to each other. Multi-instance learning has been able to produce successful results when applied to tasks such as image categorization/retrieval/annotation , text categorization, spam detection, medical diagnosis, face/object detection, object class discovery, and object tracking.
Multi-instance learning can have two different cases which are heterogenous cases (each instance is categorized by a different rule), and homogenous cases (each instance is classified by the same rule. Heterogenous cases are much more difficult for multi-instance machine learning algorithms to successfully solve.
Inaccurate weak supervision is based on training data sets which some of its label information is known to have errors or not always be representing ground-truth. Inaccurate weak supervision seeks to solve the problem of inaccurate data by successfully identify the potential mislabelled data and make the necessary corrections.
A notable scenario related to inaccurate weak supervision occurs with machine learning crowdsourcing work done by organizations such as Amazon Mechanical Turk (AMT). AMT contracts workers to label data based on their own judgements and the accuracy of those labels vary based on the labelling ability of each worker. Inaccurate weak supervision seeks to correct the labelling errors produced by crowdsourcing the labelling processes of machine learning data sets.
Time-Delayed supervision is a method of reinforcement learning which can be considered to be a type of weak supervision.
