A machine like Chandra can learn through a supervised learning process of the type attempt -> feedback -> new attempt. First, the algorithm tries to recognize an object and give it a name. During the learning process, the data scientist compares the “real” name of the object with the “attempted” by the algorithm: if the two do not match, the correct answer is provided to the mathematical model, which takes advantage of this correction to be more accurate when a new object is presented to it.

When an adequate number of defective parts are not available, it can be difficult to apply supervised training methods. In these cases, a frequently applied strategy is to use anomaly detection methods to identify rare objects or details that raise suspicions by significantly differing from most other images.

To create a performing algorithm it is necessary to have consistent, good quality data. Even in the presence of a limited amount of real data, however, a robust system can be achieved thanks to synthetic data or artificial data that simulate real phenomena. By applying suitable ‘transformations’ to the available data, it is possible to create new data to achieve a wider set of realistic situations. Moreover, Transfer Learning techniques allow the transfer of learned knowledge to solve a similar, but not identical problem, to a problem on which the algorithms have already been trained.