Flexible FFT design in deep CNNs.
During this seminar, Han will present the results of his postdoctoral research.
Deep Convolutional Neural Networks (CNN) widely employed in computer vision and the modern AI systems such as image classification and object detection, achieving near-human performance capabilities. However, the bottlenecks of the CNN hardware design are typically very computation and memory intensive, making them difficult to deploy on embedded systems with limited hardware resource such as smartphones or ubiquitous electronics for the internet-of-things. To tackle this limitation, this framework presents hardware implementation of the flexible and configurable Fast Fourier Transform (FFT) in deep CNNs.