Referent:

James Knight, Advanced Processor Technologies Research Group, University

of Manchester, UK

Abstract:

Convolutional neural networks (ConvNets) are inspired by the organization of the mammalian visual cortex and are currently the best performing solution to a large range of problems ranging from speech synthesis to image classification. In this talk, I will first compare the structure of ConvNets to that of the type of more realistic cortical models employed by theoretical neuroscientists. I will then discuss how the structures of these models affect efficiency when running them on different computer architectures: whereas GPUs are well suited to both training and performing inference using ConvNets, supercomputers or specialized neuromorphic hardware are currently the best means of running large-scale models of the cortex.

In the final section of this talk, I will discuss recent developments in bringing these two classes of models closer together. These developments include transferring ConvNets trained on GPUs to neuromorphic hardware – allowing inference to be performed at low power using spiking neurons.

Zeit:

Wednesday, 23 November 2016, 15:00

Ort:

Jülich Supercomputing Centre, Rotunda, building 16.4, room 301

Ankündigung als pdf-Datei: