Optimizing convolutional neural networks for embedded platforms
It is widely known that neural networks applications are computational and memory intensive. Apart from the atomic computations of layers (convolutions, activations, fully connected layers, etc.) in the convolutional neural networks (CNNs), the base architecture of the CNN used in the inference is prominent in embedded platforms. Present ecosystem of the embedded platforms with respect to processing of CNNs are not completely matured yet, so as to be used as out-of-box solutions in a fully autonomous vehicles. Thereby, optimization of CNN is imperative for any embedded platform. This talk will first visit on various techniques in linear algebra that can help in achieving required reduction in computation and memory footprint. Then the various techniques that can actually reduce computational complexity of the CNNs. The talk will be focused on the optimization aspects of CNN for computer vision applications targeted for fully autonomous driving.
Pallab Maji is senior research engineer for autonomous driving at Mercedes-Benz Research and Development India at Bangalore. He leads a team of machine learning research engineers at MBRDI to work on interaction of autonomous vehicles with humans in the road through computer vision. He received masters’ and doctorate degree in electronics and communication engineering from National Institute of Technology, Rourkela India. His primary focus of research was on developing artificial intelligence algorithms like fuzzy logic and neural networks for various platforms like DSPs and FPGAs. His research interests includes machine learning, computer vision and embedded platforms.