Efficient implementation of compact deep neural networks for mobile applications

Computer vision has become part of our daily life with applications in surveillance cameras, medical imaging, self-driving cars, and so on. Recently, deep convolutional neural networks (CNNs) have shown substantial progress in many vision applications such as image recognition and object detection. Utilizing CNNs from 2012 have reduced the error of image recognition dramatically which is now below than human level (i.e. 2.75% vs. 5%). Deep CNNs, however, require a large number of parameters and computations which make them difficult to deploy into smart mobile devices. Several studies have been conducted to find solutions to reduce the complexity of CNNs. Although they showed promising results for some benchmarks, the generalization of the methods into real-world applications has not been fully understood yet. In this work, we aim at finding the best method to compact deep CNNs and to embed them into cell-phone platforms.

Faculty Supervisor:

Pierre Langlois

Student:

Partner:

Huawei Technologies Canada Co Ltd (Montreal, QC)

Discipline:

Computer science

Sector:

Technology; New and Digital Media; Other

University:

École Polytechnique de Montréal

Program: