Julian Höfer, M.Sc.

Julian Höfer, M.Sc.

  • Engesserstr. 5
    76131 Karlsruhe

Research interests

Reliable AI models and accelerators in safety-critical environments

The basis for autonomous driving and other safety-critical applications is reliable perception of the near environment using cameras, as well as radar and lidar sensors. Machine learning, e.g. Convolutional Neural Networks, provide the best results for object detection. The current challenge is to integrate these neural networks into embedded systems while ensuring reliability. In particular, random hardware faults in AI accelerators as well as the inability of neural networks to estimate their own uncertainty still prevent safety-critical deployment.

AI accelerator - algorithm co-design and design space exploration

With the growing number of applications for machine learning, the requirements not only increase in terms of algorithmic accuracy, but also in terms of hardware complexity. Important goals are the minimization of memory demands and the reduction of power consumption. Optimizations are possible not only on the algorithmic level but also on the hardware architecture level. The trade-offs must be mutually weighed for the respective application to find the best solution. Co-design methods and the targeted exploration of the design space achieve the best results.

Energy-efficient AI hardware accelerators

Nowadays, machine learning can already solve complex problems in image processing for autonomous driving, industrial automation or fault detection surprisingly well. A disadvantage of such systems still remains the high computational effort and associated power requirements. For this reason, both in research and in industry (Google, Tesla, ...), special hardware architectures are being developed to implement the algorithms efficiently. At ITIV, we are working on new concepts and ideas for hardware acceleration in the field of machine learning and artificial intelligence.

Student works
title type