Fast Training of Convolutional Neural Networks via Kernel Rescaling

Talk Abstract:

Training deep Convolutional Neural Networks (CNN) is a time-consuming task that may take days or even weeks to complete. In this talk, I will present a method for training CNNs that exploits the spatial scaling property of convolutions to reduce training times without incurring any loss in accuracy. When applied to architectures such as ResNet and large datasets such as ImageNet, this method can a reduce training times by nearly 20%.  

Speaker's Bio:

Pedro Porto Buarque de Gusmão received his BSc degree in Telecommunication Engineering from the University of São Paulo in 2010. As part of a double degree program, he obtained his MSc degree in the same field from the Politecnico di Torino-Italy in 2009. During this period, he was also a recipient of an Alta Scuola Politecnica scholarship and did a one-year internship in Telecom Italia Labs. He obtained his PhD from the same Politecnico in 2017, where his main research topics involved feature-based navigation fast training of CNN.