Back to ASAM International Conference 2019

Convolutional Tunneling and Its Application in Neural Network

Intakhab Khan
Founder & Managing Director
Automotive Artificial Intelligence (AAI) GmbH 

co-authors:

Bo Tian  |  Yuanbo Xiang

Automotive Artificial Intelligence (AAI) GmbH 

For autonomous driving, it is common to use neural networks to model perception and derive decisions. Problematic is that the available data for training is usually limited and thus the training result is prone to be over-fitted. This phenomenon is also referred to as lack of generalization, which is dangerous since rare or extreme cases are not predicted correctly. This paper provides a way to train better generalized neural networks based on sparse data but still avoiding the danger of over-fitting. The findings for this paper were derived during research for developing human-like intelligent traffic agents for realistic traffic Simulation integrated into a Simulation tool used for testing and Validation of automated driving algorithms. In this specific case, a supervised learning approach based on real driving data is used to derive a human-like decision maker for the traffic agents.