Machine learning refers to various methods for deriving patterns from data that can be used to interpret new inputs. This is what is meant by computers that can "learn" without being explicitly programmed. Machine learning techniques are often associated with the development of artificial intelligence, but they are also prevalent in applications that involve large amounts of data, such as in image and speech recognition, and in risk assessment for business and finance.
Broadly speaking, there are three forms of machine learning: (i) supervised, in which a machine infers a function from a sample of input-output relations (called training data); (ii) unsupervised, in which a machine tries to uncover a hidden structure from assorted data; and (iii) reinforced, in which a machine attempts to develop a strategy for winning a game given a set of rules and objectives.
Formally, machine learning techniques involve data represented as vectors in a high-dimensional space. As it turns out, quantum information has taught us that quantum computers are good at manipulating high-dimensional vectors for certain tasks. The general aim of quantum machine learning is to develop quantum algorithms that exhibit significant speedup over classical machine learning techniques.
In many cases, the quantum learning approach seems to amount to running a classical learning problem on a quantum computer, hoping to find a speedup. But in other cases such as neural networks or Bayesian decision theory, the approach is to translate stochastic methods into the language of quantum information, hoping to develop purely quantum methods with no real classical counterpart.
