Can Knowledge Transfer in a Genetic Algorithm Accelerate Hyper-Parameter Tuning of a Deep Neural Network?

Solid State of Mind uses automatic methods to adjust the details of the architectures and learning methods (i.e., hyperparameters) for some of its algorithms. However, this process is often slow and costly in terms of resources and energy. Through this internship proposal, Solid State of Mind wishes to study the possibility of accelerating the automatic adjustment of these hyperparameters through a knowledge transfer process. The proposed project consists in developing a genetic algorithm allowing not only the evolutionary transfer of the hyperparameters of an artificial neural network from one generation to the next, but also of the knowledge acquired by the preceding generation to the next. In other words, not only transferring innate structure through reproduction, but also having the parents teach their kids what they know before the kids leave the nest to learn on their own and try to best their parents.

Faculty Supervisor:

Ioannis Mitliagkas

Student:

Partner:

Solid State of Mind Inc

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

Université de Montréal

Program: