عنوان سخنراني: Binary quantization of deep neural networks using regularization
زمان برگزاري: شنبه 21 ارديبهشت ماه 1398
Title: Binary quantization of deep neural networks using regularization
Abstract: The deployment of deep neural networks on edge devices such as cell phones has been diffr cult because they are resource hungry. Binary neural networks help to alleviate the prohibitive resource requirements of deep nets, where activation function and weight are constrained to only one bit. We modify the back-propagation, by introducing a regularization function that encourages training weights around binary values. Our simple modifr cation beats the state of the art binary deep networks in complex image classification tasks.
Speaker: Dr. Vahid Partovi Nia
Principal Machine Learning Scientist at Noah's Ark Research Lab of Huawei Technologies, Adjunct Professor at Ecole Polytechnique de Montreal