7. AI Systems Tübingen Meetup

Name: 7. AI Systems Tübingen Meetup
Start: 2023-05-04T18:00:00+02:00
End: 2023-05-04T20:00:00+02:00
Location: Sand 14

Hosted By

Oliver B. and 4 others

Details

We are happy to announce the 7th AI Systems Tübingen Meetup hosted by the University of Tübingen and IBM Research & Development. This time, we prepared two interesting talks for you:

Optimized Training and Deployment for Edge AI (Christoph Gerum, University of Tübingen)
Hardware-aware neural architecture search and co-design of machine-learning applications and hardware accelerators are essential building block for the devolopment of embedded devices such as smartphones, IoT devices, and drones. By taking into account the hardware constraints during the design process, one can create more accurate, efficient, and power-aware systems that can operate in real-time with minimal energy consumption.
This talk introduces the neural architecture training and deloyment framework (HANNAH) developed at the embedded systems group at the University of Tübingen. And shows its application to the development of Edge AI Applications.
Using Reinforcement Learning for Repeater Tree Optimization (Gregor Boronowsky, IBM R&D Böblingen)
Buffering is a crucial aspect of developing integrated digital circuits, where buffers are strategically inserted into the chip's routes to minimize propagation delay, power consumption, and noise. It is necessary to meet the strict timing requirements of modern high-performance processor designs. However, classical algorithms are often unable to produce optimal buffer trees, leading to a considerable amount of manual effort.
This talk presents "aiBuf", an AI agent that employs reinforcement learning to autonomously optimize buffering. By interacting with the environment and receiving feedback in the form of rewards, aiBuf can find more efficient and effective solutions than classical algorithms. The use of aiBuf can significantly reduce the manual effort required in buffer optimization, with improvements of up to 85%.