Enable AI models to run on the edge

Purpose and goal:

This project will determine if two of Volvos AI use cases can be deployed to Volvo´s existing as well as future hardware platforms. The project will demonstrate potential performance improvements for the selected Volvo use cases on selected embedded hardware targets using Embedl´s state-of-the-art optimization methods. In addition, it will show how Embedl contributes to improving productivity and agility.

Expected results and effects: A better understanding of how to use deep neural networks for in-vehicle applications including: 1. Network compression techniques for faster inference 2. Adjusting the training processs to be aware of post-training optimization 3. A recipe for choosing appropriate hardware for inference

Approach and implementation: The work will be carried out in 4 work packages with the following distribution of responsibilities: WP1 - Setup (resp. Embedl + Volvo) set up the dev/test environment for running models provided by Volvo on target HW WP2 - Baseline (resp. Embedl) where the baseline performance of models provided by Volvo is measured on target HW

WP3 - Improve (res. Embedl) apply Embedl´s optimization techniques to improve model inference

WP4 - Verify (resp. Embedl + Volvo) infer optimized models on target HW and measure if gains in inference throughput and regression in performance is acceptable