Training of Neural Networks via Gradient Descent and Solvers

In Deep Learning, one of the most important determinants of the performance of a model is the training procedure used. The ideal training method produces a model that is accurate and fast, it should work on models of various sizes and be flexible enough to allow users to specify conditions they want the model to satisfy. This proposal aims to create a DL training algorithm that achieves all of these by using a careful blend of two different methods of training. The two different methods in question -gradient descent and mixed integer linear programming, each have their own strengths and limitations, this work will combine the two in an attempt to create a superior “hybrid system”. A better training algorithm has simple yet powerful applications in the real world. Every single DL system – from financial risk prediction to data analytics would be more accurate, more customizable and all this could be achieved with fewer resources.

Faculty Supervisor:

Sheila McIlraith;Vijay Ganesh

Student:

Partner:

Royal Bank of Canada (Borealis)

Discipline:

Computer science

Sector:

Information and Communications Technology; Technology; Finance and Insurance

University:

University of Toronto

Program: