Machine learning algorithms process data in order to build mathematical models that in turn can be used
to help make predictions and/or decisions. These algorithms are not explicitly programmed with
instructions for how to solve a problem. Instead, they improve autonomously (or ‘learn’) from experience.
The systems learn to generalise from example data, with minimal human intervention.
Machine learning is particularly useful when working with very large datasets. It is well suited to tasks that require repetitive routine activity (such as interpreting scans) and, for some tasks, can perform faster and more accurately than a human interpreter.
There are many different types of machine learning algorithm, and the most appropriate algorithm
depends upon the specific nature of the task at hand. Here, we briefly summarise three categories:
reinforcement learning algorithms.
Supervised learning algorithms perform a complex ‘connect-the-dots’ operation between a given set of input data and an associated output.
They are trained with multiple examples of a set of inputs and a paired known output, learning how to process the inputs in order to
reproduce the related output. The fully trained algorithm can then be given novel sets of inputs, for which the outcome may not be known,
and make a prediction as to what the related output should be.
For example, an algorithm may be trained to classify whether the configuration of pixels in a picture (the input) represents an image of an apple (the output).
Supervised algorithms are used for two types of problem:
classification (to predict which ‘class’ an observation belongs to, eg, case vs control), and
regression (to predict a continuous value, eg, time to diagnosis).
Unsupervised learning algorithms lack the ‘supervision’ of a known set of output information.
They instead process input information to identify consistent patterns and associations between
variables. The output they produce is a grouping or summary of the input data, rather than a targeted
Reinforcement learning algorithms are concerned with producing an action depending upon a given state.
This ‘state’ refers to a configuration of inputs or observations from the environment the algorithm is operating in.
They process a given state, decide upon an appropriate action based upon the state, observe the consequences of that action, and then
‘reinforce’ or penalise the state–action pairing accordingly. This makes the same action more or less likely for future
scenarios in which a similar state is encountered. With multiple cycles of the same process, actions can be produced that maximise a
rewarding outcome and minimise risk. The slow iterative nature of this process makes it well suited to tasks where there may be
inconsistent or limited outcome events. For example, reinforcement learning algorithms have been used to great effect in optimising
algorithms for playing chess and other board or computer games, where the ‘reward’ signal is fairly sparse or delayed that is, actions
eventually lead to winning or losing the game, rather than receiving immediate feedback.