The technological challenge in the application of supervised learning becomes more and more as the data generated by society and businesses increase exponentially, and it becomes harder to build models from small training sets. Labelling process becomes a major blocking issue with limited number of subject-matter experts to label them.
Reinforcement learning is a solution where AI can test the outputs in the real world, by iterative test and improvement. A very well known recent application of this technique has been implemented in the AlphaGo project.
However, there are many cases where the afore-mentioned process is not feasible. Referring to the earlier example, it is impossible to ask doctors the whole set of label millions of sample MRI scans.
For the big data problems where traditional approaches become insufficient, Active Learning (AL), which can be defined by learning by asking questions, come into the picture.
As an analogy to children learning process, supervised learning can be viewed as learning with observations, reinforcement learning can be considered to be trial and error, and AL can be deemed to be learning from a teacher by asking questions.
To make AI a viable alternative, or even a major supporting element to the human decision process, the learning process of AI needs to be enhanced. To accomplish that goal, the question process should become efficient and intelligent. AI should be able to learn the maximum amount of information with a minimum number of questions, which is proportional to experts’ time and money spent on the learning process.
Active Learning is about AI finding the right questions to learn.
Artificial Intelligence (AI) systems, just like any intelligent system, needs to learn to produce a valuable outcome. In today’s applications, the value of the result is usually defined by correctly predicting a specific label for an unlabeled set item.
Many AI systems rely on supervised learning, where AI utilizes a group of labelled samples to build a model, which later can be applied to estimate the unlabeled samples. A simple example being AI trying to diagnose undiagnosed MRI scans from the information it has learned from a diagnosed set.