Equivariant Siamese Neural Networks

The world we live in is ripe with symmetry. From the bilateral symmetry we see in humans to the symmetries which are used to describe fundamental particles in physics. Most modern machine learning methods however do not have an inherent modeling of symmetry in them. By developing algorithms which do have an explicit modeling of symmetry we can decrease the amount we need to teach these algorithms, making them much cheaper to create. We propose a network that can compare images in such a way that it is not affected by changing the orientation of objects in the image. This is useful for things such as facial recognition, where it could be used to verify a face regardless of its orientation. The purpose for this project is to create a method for robust object comparison so that it may be used for industrial applications.

Faculty Supervisor:

Nicholas Touikan;Moulay Akhloufi

Student:

Max Hennick

Partner:

New Brunswick Research and Productivity Council

Discipline:

Statistics / Actuarial sciences

Sector:

University:

Program:

Accelerate