COMPSAC 23: Presenting acceleration solutions based on Deep Neural Networks (DNNs) for use in safety-critical systems

Date: June 29, 2023

BSC researcher Martí Caro presented “Efficient Diverse Redundant DNNs for Autonomous Driving” on 27 June 2023 at the Autonomous Systems Symposium (ASYS) within the 47th IEEE International Conference on Computers, Software & Applications (COMPSAC).

The theme of COMPSAC 2023 was Resilient Computing and Computing for Resilience in a Sustainable Cyber-Physical World. It was held 26-30 June 2023 in Torino, Italy. COMPSAC, a central international forum for nearly 50 years, brings together academia, industry, and government to present research results and advancements, discuss emerging challenges, and explore future trends in computer hardware, software technologies, systems, and applications.

Within COMPSAC 2023, the ASYS served as a forum for discussion of ideas and results in a wide spectrum of topics related to autonomy in technical systems covering theory, design, implementation, application and analysis of autonomous systems. Autonomous behavior is defined as an ability to act without direct supervision from outside entities (humans or other devices). Such behavior, recently implemented in many practically utilized systems such as autonomous vehicles, autonomous software agents, autonomous robots and others, is becoming prominent for the next generation engineering systems. The symposium fostered highly fruitful and constructive discussions involving top-level experts both from academia and industry about the current state and future perspectives of autonomous systems.

Visit the event page for more information about SAFEXPLAIN participation in ASYS.

Marti Caro presented an initial research contribution towards making acceleration solutions based on Artificial Intelligence (AI), and particularly those based on Deep Neural Networks (DNNs), amenable for their use in safety-critical systems, specifically for automotive systems.

He presented the TRUST heuristic that seeks to build diverse redundant accelerators based on the use of lower precision arithmetic to reduce costs while preserving performance and reusing data fetched by the primary accelerator

  • This strategy provides effective error correction with 3-6% energy reductions with regard to DCLS-like solutions
  • Future work will focus on employing it in an actual diverse and redundant accelerator by exploiting the paper´s findings and evaluating the TRUST heuristic

Want to know more? The slides presented at the conference are available here.