Artificial Intelligence (AI)

The Technology

Artificial Intelligence (AI) made major breakthroughs in the past decades and multiple industries now rely on these new technologies to revolutionize various complex tasks and help workers improve their efficiency. Among various technologies tagged “artificial intelligence”, a subset is referred to as “machine learning”.

A machine learning model is nothing more than a mathematical model representing the relationship between a set of input parameters and an output result. The particularity of machine learning is that the model is “learned” from experimental data using an automatic learning process and statistical methods.

The learning process may be implemented in many ways, but the most common strategy is called “supervised learning” which requires the creation of a training database containing many examples of the relationship between the machine learning model inputs and the expected output results. During the “learning” or “training” process, the machine learning model parameters are adjusted or optimized to reproduce the inputs-output relationship on the examples of the training database. If the database is carefully crafted and the model is chosen properly, one may expect that the training model will not only be able to accurately reproduce the inputs-output relationship on the training database, but also produce correct results on new data that is not part of the database: the model is said to “generalize” to new data.

The selection of the training data is very important to ensure that the training process delivers a model with good generalization properties. In particular, the database must be representative of the expected data in real life scenarios. For example, if we want to develop a model to recognize cars in a picture, we need to build a database with pictures of all common car shapes, such as 4-door sedans, pick-up trucks, and SUVs. If only sedans are included in the database, the model will not be able to generalize to the other car shapes. To correct this situation, more pick-ups and SUV examples must be added to the database, and the model must be retrained. This is typically a continuous process; if the retrained model has difficulty on other car shapes, new examples are added to the training dataset and the model is retrained again, and so on. For these reasons, machine learning models are expected to improve over time if the training database is continuously expanded with new, well-chosen examples.