Attention in Convolutional Neural Networks
Convolutional Neural Networks (CNN) have shown remarkable classification performances on many different datasets. However, these networks need couple of days to train for example on ImageNet and even during inference the processing time of one image takes too long to run state-of-the-art networks in real-time, especially in context of a driving-assistant system.
In this discussion group I would like to discuss possible ways to either speed-up inference or reduce the computational load of the network by introducing attention-mechanisms.
The starting point is the 'communication-through-coherence' hypothesis proposed by Fries [Fries 2005, Fries 2015, Buschman and Kastner 2015].
|Tue, 03.05.2016||14:00 - 16:00||Sala Panorama|