Viric Learning - A Novel Transfer Learning Method

Abstract

Increasingly, information systems include intelligent components to make decisions or perform actions based on data, owing to advances in data science and artificial intelligence. Although advances have been made at the machine level to enable the execution of more complex models, due to the changing context they can quickly become outdated and require a high number of resources to adapt to new conditions, thus transfer learning is being investigated. This discipline seeks to optimize learning tasks of an intelligent model based on the learning from another model. Therefore, a method is proposed that focuses on what all models receive in a generic way: data. The objective is to elaborate a design data register that, in controlled training sequences, transfers only the quality knowledge of the model that from it was created to a new one by means of optimization algorithms. Likewise, it will also obtain clarity on the criteria used by the models to perform their functions, the development of objective mechanisms to obtain traceability of the training data and an optimal method of adaptation to the changing context to which the information systems are subjected.