Encouraging Validatable Features in DNNs

This repository contains the code to reproduce the results found in the paper "Encouraging Validatable Features in Machine Learning-based Highly Automated Driving Functionss" by O. De Candido, M. Koller, and W. Utschick, published in the IEEE Transactions on Intelligent Vehicles in 2022.

How to use the code

1. Store the training data in data/ folder, e.g., the extracted lane changes from the highD dataset from this project.
2. Run main.py to train the DNNs with the network architectures stored in parameters/params.py.
3. Run extract_embeddings.py to extract the feature embeddings from the trained networks, and calculate k-means on those embeddings.
4. Run umap_embeddings.py to calculate the UMAP representations of the feature embeddings.
5. (Optional) Update the network architectures in parameters/params.py and rerun the scripts.

Requirements

The required packages can be installed via the requirements.txt file, e.g., conda install -r requirements.txt. This code was tested using Python v3.7.0.

Paper Reference

@article{decandido2022encouraging,
    author={De Candido, Oliver and Koller, Michael and Utschick, Wolfgang},
    title={Encouraging Validatable Features in Machine Learning-based Highly Automated Driving Functions},
    journal={IEEE Trans. on Intell. Vehicles},
    year={2022},
    doi={10.1109/TIV.2022.3171215},}
}

Note

This paper is an extention of our 2020 IEEE ITSC paper. The code relating to that project can be found here.