Making the most of data with weak supervision
In between completely supervised and unsupervised learning are several approaches to partial supervision, where only x data is supervised. Like unsupervised approaches, the goal of these techniques is to make the most of supervised data, which can be expensive to obtain. One advantage of partial supervision over unsupervised approaches is that unsupervised results don’t automatically have useful labels. The labels have to be supplied, either manually or through some of the techniques we saw earlier in this chapter. In general, with weak supervision, the labels are supplied based on the subset of the data that is supervised.
This is an active research area, and we will not go into it in detail. However, it is useful to know what the general tactics of weak supervision are so that you will be able to apply them as they relate to specific tasks, depending on the kind of labeled data that is available.
