Description

Machine Learning denotes the field of algorithms enabling computers to learn and improve at a specific task without being explicitly programmed.

This field is very wide, and makes use of several scientific disciplines, especially related to statistics and optimization.

Machine Learning algorithms can be divided in 3 main categories: supervized, unsupervized, and reinforcement learning.

Supervized learning

Based on a set of labelled (or categorized) data points, an algorithm is used to create a model (mathematical representation; regression or classification) of this set, so as to be able to predict as accuretaly as possible the label of new data points. For instance, an algorithm can be trained to distinguish pictures of airplanes and helicopters (from a training set of numerous such pictures, correctly labelled). Then, it should be able to determine if a picture it was not trained on is that of an airplane or an helicopter.

Classic examples of such algorithms are Naive Bayes classifier, Support Vector Machine, Artificial Neural Network, Random Forest.

Unsupervized learning

Unlike supervized learning, the set of data points is not labelled. The aim in this case is to build a model based on solely on the information contained in the set, i.e. the commonalities in the data.

The main approaches to do so are clustering (find groups of similar data points; e.g. k-means, hierarchical clustering), neural networks (e.g. Generative Adversarial Networks, autoencoders) and extraction of latent variables (e.g. Principal Component Analysis, Singular Value Decomposition).

Reinforcement learning

A software agent in its environment (e.g. game AI, aircraft autopilot) is trained based simply on the consequences of its act, so that its actions maximize its reward (e.g. game score, trajectory compliance).

Typical methods to achieve this result include dynamic programming with Markov Decision Processes, online (e.g. Q-learning) and offline (e.g. Least Square Policy Iteration) reinforcement algorithms, and simulation methods (e.g. Monte Carlo Tree Search).

Various examples

Machine Learning applications are very diverse. Following are a few examples of these.

Image processing (extracting information from images) can be used to detect people (see facial recognition) or cars on surveillance camera, help doctors make diagnoses based on medical images, inspect parts during production or maintenance, ...

Another main field of application is prediction: predictive maintenance (determining when a part is likely to fail, in order to replace it just before it happens), fraud detection (detecting ominous patterns), product suggestion on retailing websites, ...

Machine learning is also the technology behind chatbots, powering speech recognition engines.

Key references

• Andrew Ng (Stanford): Machine Learning course
• Kaggle: data science competitions (examples of application)
• IATA: AI White Paper