227-0085-29L Projects & Seminars: Practical Embedded Deep Neural Networks with Special Hardware Accelerator
|Autumn Semester 2020
|every semester recurring course
|Language of instruction
|Only for Electrical Engineering and Information Technology BSc.
The course unit can only be taken once. Repeated enrollment in a later semester is not creditable.
|The category of "Laboratory Courses, Projects, Seminars" includes courses and laboratories in various formats designed to impart practical knowledge and skills. Moreover, these classes encourage independent experimentation and design, allow for explorative learning and teach the methodology of project work.
|Deep neural networks (DNNs) have become the leading method for a wide range of data analytics tasks, after a series of major victories at the ImageNet Large Scale Visual Recognition Challenge (ILSVRC). For ILSVRC, the task was to classify images into 1000 different classes, many of which are difficult to distinguish (e.g. many classes are different breeds of dogs). All that was given were 1.2 million labelled images. Meanwhile, this recipe for success has taken over many more areas, from image-based tasks like segmenting objects in images, detecting objects, enhancing images using super-resolution and compression artifact reduction, to robotics and reinforcement learning, and a wide range of industrial applications.
DNNs and their subtype convolutional neural networks (CNNs) have not been new in the 2013 when the wave of success has started, but they got this huge boost through the new availability of large-scale dataset and—at least as importantly—the availability of the necessary compute resources by using GPUs to perform the computations required during training.
While GPUs were then also used to stem the high computation effort of DNNs during inference (e.g. classifying images directly using a trained DNN rather than training the DNN itself). The high demand, the need for cost efficiency, and the goal of deploying DNNs not just in data centers but pervasively in everyday devices, wearables, and low-latency industrial or interactive applications, has triggered the development of various application-specific processors which are much faster, vastly more energy efficient, and cheaper at the same time—such as the Google TPU, Graphcore, …, and Huawei’s Ascend/Atlas platforms.
In this course, you will learn:
1) the basics of deep neural networks, how they work, and what challenges there are for inference,
2) how platforms with specialized hardware accelerators, specifically the Huawei Atlas 200, can be used for running DNN inference and getting a practical application running, and
3) work on your own project using DNNs and hardware accelerators based on your own ideas or on some of our proposals.
The course will be taught in English by the new D-ITET center for Project-Based Learning and a special guest lecturer from Huawei. Individual interactions/help can also be in (Swiss) German.
Most sessions will be around 1 hour of lecture and 2 hours of practical computer exercises. We will start an introduction and then you will have ca. 8 weeks to work on your project, which will concluded with a final presentation of your results.