Titre : Training Binary Deep Networks

Conférencier : Vahid Partovi Nia

Résumé :
Deep neural networks (DNN) are widely used in many modern applications, such as vision, speech, and natural language understanding. However, deployment of DNNs on edge devices has been challenging, because they are resource hungry. Binary networks (BNN) help to alleviate the prohibitive resource requirements of DNN; where both activations and weights are limited to one bit. Training deep networks is difficult, but training binary deep networks is even more challenging. Here I devise a regularization function to add to the objective function, in order to encourage binary weight and activation. This approach helps to reduce accuracy degradation to a great extent. This method is based on linear operations that are easily implementable into the training framework.

Bio :
Vahid Partovi Nia is Principal Machine Learning Scientist at Noah's Ark Research Lab of Huawei Technologies in Montreal. He is a member of the Canada Excellence Research Chair in Data Science and Adjunct Professor at Ecole Polytechnique de Montreal. His mission is to make a bridge between the expertise in academia, and the cutting-edge technology in the domain of machine learning and artificial intelligence in industry. His research interest is diverse and is application-oriented including Data Science, Efficient Computing, Biostatistics, and Artificial Intelligence. He was Swiss SNF funded postdoctoral fellow at McGill University and Stanford University, and received his Ph.D. in Statistics from Ecole Polytechnique Federale de Lasuanne, Switzerland.

Bienvenue à tous!