Master Thesis Accelerating Winograd Convolutions on Edge (f/m/x)
SOME IT WORKS. SOME CHANGES WHAT'S POSSIBLE.
SHARE YOUR PASSION.
More than 90% of automotive innovations are based on electronics and software. That's why creative freedom and lateral thinking are so important in the pursuit of truly novel solutions. That’s why our experts will treat you as part of the team from day one, encourage you to bring your own ideas to the table – and give you the opportunity to really show what you can do.
We, the BMW Group, offer you an interesting and varied Master thesis in the area of Hardware/Software Codesign.
Convolutional neural networks (CNNs) have found their way into a wide range of computer vision tasks like image classification, object detection, and semantic segmentation. However, several challenges arise when deploying resource-intensive CNNs on resource-constrained embedded hardware. The computational and memory demands of CNNs are often mitigated by resorting to CNN compression techniques, fast convolutional algorithms and/or dedicated hardware accelerators. One such fast convolutional algorithm is the Winograd algorithm that has demonstrated to lower the arithmetic complexity by turning convolutions to element wise matrix multiplications, saving substantial time and energy. The goal of this thesis is to efficiently implement quantized Winograd convolutions on edge hardware with minimum performance degradation on state-of-the-art CNN workloads.
What awaits you?
Literature survey of the state-of-the-art CNN acceleration methods for edge hardware.
Experience in implementing advanced convolution algorithms and solving the challenges associated with quantization of operands.
Engagement in a diverse team with experience in publishing at international peer-reviewed conferences.
Scientific writing of your thesis and presentation of the research results both at the university and industry.
Please note that your thesis must be supervised by a university on your part.
What should you bring along?
Currently pursuing a master's degree in electrical engineering, computer science or a comparable qualification.
Very good programming skills in C/C++, Python as well as substantial experience with CUDA programming.
Practical knowledge of deep learning accelerators and their constraints on edge .
Highly motivated and eager to collaborate in a team.
You are enthused by new technologies and an innovative environment? Apply now!
At the BMW Group, we see diversity and inclusion in all its dimensions as a strength for our teams. Equal opportunities are a particular concern for us, and the equal treatment of applicants and employees is a fundamental principle of our corporate policy. That is why our recruiting decisions are also based on personality, experience and skills.