PhD Course: Circuits, Architectures and Systems for Intelligent Computing at the Edge
Deep learning oriented to image classification, object detection and segmentation tasks has reached unprecedented popularity in several real-world applications, ranging from intelligent autonomous vehicles to smart manufacturing processes. In these contexts, ensuring real-time performances also on power-constrained devices is crucial. However, state-of-the-art deep learning models, and in particular Convolutional Neural Networks (CNNs), owe their high accuracy to massive computational and memory requirements, which hinder their deployment into edge computing systems. This issue has opened the door to an exciting study field that aims at researching new design and algorithmic strategies in order to reduce the computational complexity of deep learning models without compromising the achievable accuracy.
In this course, we will introduce the basics of CNNs with reference to different computer vision tasks. After that, we will provide an overview on techniques currently used to efficiently implement CNN inference on low-power edge devices through data-level approximations, such as quantization and pruning. Some noteworthy state-of-the-art FPGA and ASIC implementations will be also presented. Finally, we will describe more recent advances in the field of layer-level approximate computing techniques and their application to state-of-the-art models, discussing how well error-resilient tasks like CNNs can really tolerate aggressive approximations and quantifying the benefits of such strategies on power
consumptions. At the end of this course, students will have a comprehensive knowledge of the main design techniques applicable at both circuit-, architecture- and system-level to enable intelligent computing on edge devices. They also will get an understating of methodologies and tools that can be used in varying artificial intelligence applications also related to their research topics.
seminar room - 5th floor, cube 42C
05/06/2023 (10:00 - 13:00)
06/06/2023 (10:00 - 13:00)
07/06/2023 (10:00 - 13:00)
08/06/2023 (10:00 - 13:00)
12h - 3CFU