To improve traffic safety, Deep Neural Networks (DNN) are being developed worldwide for automotive applications. The challenge is that DNNs are compute- and memory-intensive, but the computing capacity in the vehicle remains limited.

Within the framework of this project methods of approximate computing are investigated in order to reduce the computational load. Approximate Computing refers to a set of methods that are performing calculations not exactly but only approximately. As a benefit, fewer resources are used in the FPGA, more functions can be implemented in the existing FPGA devices, and the energy efficiency of the calculations is improved.

We are tackling the challenge of incorporating approximate computing into the optimization process of the DNN application to reduce computational load while maintaining tolerable loss of quality. Machine learning will be used for optimizing the approximation parameters.

The project is sponsored by BMW AG for three years.