Hyper-parameter Optimization of Deep Neural Networks

Deep neural networks are a valuable machine learning method that is at the heart of many technological innovations. From self-driving cars to automatic translations and image recognition, etc. it seems that deep neural networks are a great tool that can adapt to different problematics. However, defining the right network for the right application is a tricky task, often described as a black art, that monopolizes an important part of the development process. This crucial step still relays strongly on the experience of experts, or on heuristic approaches. Our goal is to develop a rigourous scientific method to optimize the hyper-parameters of a deep neural networks, which are the elements that define the network, in order to automate this step. We also require that our method be convergente and includes a global optimization scheme.

Faculty Supervisor:

Sébastien Le Digabel

Student:

Partner:

University of California, Los Angeles

Discipline:

Mathematics

Sector:

Education; Professional, scientific and technical services

University:

École Polytechnique de Montréal

Program: