In simple terms, an algorithm is a set of instructions for converting an input into an output. Much like how the ingredients in a recipe code the process for converting ingredients into a finished meal, machine learning algorithms encode the process for converting data into specific outputs.

This is a machine learning process where inputs are identified as belonging to a specific category. This can be binary with only two potential outputs (like the spam example above), or multi-class when there are more than two categories.

A regression model is used to determine a numerical value from input data. For instance, you might want to determine how much a used car should cost on the basis of a range of factors: age, mileage, condition, make etc, giving each a different weighting. You could build these variables into a regression model, input values for each, and get a specific output such as the price you should sell the car for.

Clustering is a process where an algorithm is used to group data, without human input as to what categories the algorithm ought to be sorting by, such as in our customer segmentation example above.

One way to cluster is with a K-means algorithm. Suppose we were feeding an algorithm the height and weight of different Rhinos in a herd. Those that are heavy and taller might form a cluster, and those that weigh less and are shorter may form another. The k-means algorithm picks out an average point (centroid) for each group, and classifies data relative to its distance from this average. With our example. This might allow you to identify three different size group s(e.g. male, female and calf), without the need for a training set labelling known examples.

There are different types of layers in a machine learning process. The first, is the input layer which feeds the features into the model. For instance, if we were tracking car cost, the age, mileage and condition. The output layer is the result of the process, with the car example, the model's suggestion for an appropriate price. Between the input and the output layer can be a varying number of hidden layers, which encode further algorithms to transfer the input data to a desired output.

Each hidden layer contains a number of neurons. These are an algorithm, which contains weights and an activation function. When data is input to them from the input layer, or a previous hidden layer, they calculate whether a certain threshold has been met. If it has, they pass on an output to the next layer, mimicking the process of how information is transferred in human neurological networks. As information is transferred through these neurons (contained in each layer), the input data is processed to produce an output.

An artificial neural network describes the different layers and neurons explained to the left. We can simplify the neural network as one big algorithm, containing various algorithms encoded into the neurons that make up the network. As we increase the number of neurons and layers, we increase the complexity of the model.

A deep learning model is a neural network containing three or more hidden layers. The complexity involved in a deep learning model facilitates use-cases that go beyond simple classification and prediction to perform more complex tasks, such as voice recognition and chatbot solutions.

Machine learning models can be flexibly applied to provide diagnostics for a range of diseases. For instance, diagnosing sickle cell disease usually requires expensive laboratory equipment, highly skilled workers and reliance on transport. Gazelle, a portable, machine learning diagnostic platform can be used to detect the disease, simply, at the point of care, at a much lower low-cost than traditional methods. Developments like this can help ease the resource burden on hospitals in low-income countries and increase the number of people getting treated for diseases.

As the climate crisis worsens, and its impact on food production becomes more severe, machine learning offers a way to both increase food yields and reduce the impact of farming on the environment. Combining machine learning with sensors, can allow farmers to determine the exact amount of fertiliser and water needed to optimise production yields. Alongside this, machine learning-powered, urban farming equipment can be used to produce crops in much smaller environments that those needed for traditional farming, reducing demand for land.

Predicting earthquakes can be particularly difficult in busy cities, where seismic tremors indicating an earthquake are confused with the noise produced by bustling cities. Deep learning models have been developed to distinguish between city noise and Earthquake signals.