In 2012, Krishevsky et al, submitted an entry to the ImageNet competition that smashed the previous error rate record. Krishevsky’s solution won the competition using deep convolutional nets (conv nets) and GPUs. From 2012 the ImageNet competition error rate has continued to plummet from ~15% to just over 3%, with all winning entries using deep conv nets and GPUs. In this talk I review the history of deep conv nets, the key points from Khrishevsky’s paper, conv nets strengths and weaknesses and briefly show how easy it is to create your own deep conv net using Google’s new machine learning library TensorFlow.

Date: 03/06/2016
Time: 16:00
Location: LB252