Supervised vs UnsupervisedBoth supervised and unsupervised learning are common artificial intelligence techniques. The choice between the two is based on constraints such as availability of test data and goals of the AI.Generally speaking, unsupervised learning is used to tackle broad problem spaces such as a visual recognition AI that is expected to identity a broad range of things, events or states. Supervised learning is sufficient for well-defined problem spaces such as an industrial robot that is learning to sort parts into bins.
Supervised & UnsupervisedIt is possible to use both supervised and unsupervised learning on the same problem space. For example, supervised learning can be used to establish some base models followed by unsupervised learning. Alternatively, supervised and unsupervised learners can be used together using techniques such as deep learning and committee machines.
ExampleAn unsupervised learner analyzes images from television shows together with subtitle text. The learner identifies patterns in the images and associates those patterns with words in the text. With time the learner is able to identify objects or situations in new television shows. For example, the unsupervised learner might develop an ability to identify cars or people who are crying. The learner isn't told if its models are right or wrong but develops confidence in them with enough samples. Likewise, the learner may decide that a particular model isn't valid or useful and discard it.A supervised learner would tackle the same visual recognition tasks by viewing images that are labeled. For example, the learner could be shown images labeled as happy, sad and angry to learn to identify facial expressions.The unsupervised learner might eventually learn to identify countless objects, states and events if it were given enough processing power to watch all the television shows ever produced and study each frame. The supervised learner is constrained by its test data including expected answers.
|Supervised Learning vs Unsupervised Learning|
Machine learning based on test data with expected answers.
Machine learning based on test data without answers.