One important aspect of artificial intelligence (AI) is the ability for computers to learn from experience (a large dataset of examples) to help them improve their decision making capacity and to be able to make predictions based on this data. The software that make this possible are generally referred to as machine learning applications and are regarded as a branch of AI.

Machine learning applications are made up of algorithms that represent a sequence of statistical processing steps that are used to identify patterns and features in massive datasets that are then used to make decisions and predictions about new data. The better the algorithms, the better the predictive results of the machine learning process.

Examples of this aspect of AI learning surround us in the digital world today. Voice activated assistants, such as Alexa and Google Voice, enable applications to search the web and perform certain functions after hearing spoken instructions from a user. Websites that recommend products based on the past purchasing behaviour of a user (and other similar users) are another. Email programs now feature spam detectors that can spot unwanted emails based on their characteristics and remove them from user’s inboxes. Other examples include the ability of computers to spot tumours in patients’ xrays and aspects of the way self-driving cars assess the environment around them to drive without human intervention.

There are various methods by which AI can learn, which are referred to as machine learning styles, and fall into three general categories:

• Supervised Machine Learning

In this method algorithms operate on a labelled dataset which contains information about the data on which the model is being built to help classify the data. A computer vision application, for example, may be trained to identify certain types of dogs based on labelled images of that type of dog. This approach requires less training data than other methods because the labelled data can be compared to the results. The labelled data is, however, time-consuming to prepare and can be biased based on the training data provided.

• Unsupervised Machine Learning

In this approach unlabelled data is used in massive amounts and the algorithm must extract meaningful features of the data on its own in order to classify the data in real-time, without human intervention. This approach is less about decision making or a predictive approach and more about identifying patterns and relationships in data that humans might otherwise miss. This approach might be used to identify features and patterns of spam email, for instance, from large numbers of examples.

Semi-Supervised AI learning, this is an approach that uses aspects of both of the above methods by using a smaller labelled dataset to guide its approach in classifying new, large datasets. This approach can solve the problem of having insufficient labelled data when training a supervised learning algorithm.