The biggest area of AI research today seeks to enable computers to make
inferences from complex data. Techniques to do this are termedmachine learning. Two examples below illustrate classification (of iris flowers) and regression (estimating the quality of wine from scientific measurements).
Common to these methods is that data is represented by numerical features.
Machine learning creates a model of these features. For the two examples above, we start with some training data points where the desired output is known, and can use this to teach the model (by updating its internal weights) to predict these wanted values. Once trained, the model can then be used to make predictions with unseen data. This is known assupervised learning. However, there are many scenarios where outputs are not available, but machine learning can still be used to identify patterns in data. Models can be trained to pick out distinct, independent features, for example different types of sound, or classify data into similar groups. Other approaches teach models to predict future values in a series using historical data, for example to learn how share prices change. These processes are collectively termed unsupervised learning. Reinforcement learning is a way of getting ML to teach itself to solve problems by training to maximise a reward like a points score, and was impressively used by Deepmind to learn how to play computer games, and then the board game, Go. Model structures and training processes vary widely. Especially when data is scarce, decades-old regression methods can often give good solutions. More recent approaches include decision forests, gradient boosted decision trees, support vector machines, neural networks, belief networks, convolutional networks and recurrent networks.