Maximize Deep Neural Network Efficiency for Automated Driving
Program Director UROP
- +49 241 80-90299
- Send Email
- Project Offer-Number:
- UROP International
- Electrical Engineering, Information Technology and Computer Engineering
- Organisation unit:
- Chair of Integrated Digital Systems and Circuit Design
- Language Skills:
- Computer Skills:
- some programming experience (e.g. in Python)
- Prof. Dr.-Ing. Tobias Gemmeke
With the advancing automation of vehicles, edge computing tasks become an increasingly important research topic. These tasks require powerful hardware, which heavily influences the design process of intelligent transportation systems. Therefore, it is desirable to minimize the required energy of all computational tasks. Deep neural networks (DNNs) make up a large part of these tasks; for instance, we use YOLACT++ to process camera data. One possibility to reduce the energy consumption is the quantization of DNNs. The objective of this project is to efficiently quantize the YOLACT++ network while maintaining its baseline performance. This project will be supervised in cooperation with the Automated Driving Department of the Institute for Automotive Engineering.
Most energy is consumed by data movement; hence, the memory footprint of the DNN needs to be reduced. This project offers the possibility to explore either in PyTorch or TensorFlow to increase the energy efficiency of YOLACT++. - Quantize the current floating point architecture with minimal impact on performance. - Experiment with further methods for reduction of the total memory footprint, e.g., pruning.
Programming experience, interest in Python programming, interest in Deep Neural Networks and frameworks like PyTorch or TensorFlow