Project

Our starting position and challenges

So far, AI modules in autonomous driving applications are scalable to a limited extent only. In fact, they only react reliably in limited scenarios. The reason lies in the training strategies applied, in which algorithms are retrained for each new domain. This results in enormous development costs.

KI Delta Learning researches disruptive methods to train the AI algorithms in a more sustainable way. New, more efficient training methods ensure continuous learning. Knowledge that has already been learned as well as previously tested and secured development levels are retained when changing domains.

This is the only solution to keep up with ever shorter innovation cycles and constantly changing mobility. Current technological gaps that limit the Technology Readiness Level (TRL) of autonomous vehicles and slow down a broad application of AI in autonomous driving will be closed by new approaches in KI Delta Learning.

Our objective: Autonomy at scale

AI used in autonomous vehicles must be responsive to a constantly evolving market and scalable to meet changing requirements. Typical examples of domain changes – deltas - are different sensors as well as changes in time and location.

Different domains of AI in automotive application

Project objectives – bridging deltas

To remedy these deltas, the project focuses on three main areas for delta learning: transfer learning, didactics and automotive suitability.

The state of the art is being advanced in all three areas to such extent that the next generation of AI algorithms will be suitable for an unrestricted use in autonomous vehicles.

Furthermore, a project-specific data set, specifically tailored to the project objectives, will provide the basis for developments in these three areas.

Transfer learning: the delta between previously trained and new domains

New tools and methods enable the AI modules to be extended and transferred to new domains and new tasks without significantly reducing performance in the original areas.

Didactics: the delta in the learning process

Not only the knowledge itself, but also how the knowledge is conveyed, has an influence on the learning results. Therefore, the focus should be on targeted learning strategies and model structures for effective and efficient training.

Automotive suitability: the delta between requirements in the automotive industry and current AI approaches

AI modules must have important properties such as robustness and real-time capability in the automotive context. New domains due to changes in the open world represent deltas in addition to new hardware requirements for integration into the embedded environment.

The project’s own data set: the delta between general training data sets and tailor-made data

Freely available training data sets are often unspecific and only of limited use for achieving the project objectives. The collection of a comprehensive project-specific data set provides for the first time the basis for an efficient development of delta learning methods.