Bosch Center for Artificial Intelligence
Our Research Field

Rich and Explainable Deep Learning

What Motivates Us

Perception is a key enabler for driver assistance and autonomous cars. Today, deep learning is the state-of-the-art tool for perception providing the highest accuracy. However, while there has been significant progress using deep learning approaches, further advancements are required to achieve sufficient perception accuracy and robustness based on various sensors. In addition, establishing increased reliability is crucial for deployment of artificial intelligence algorithms in safety-critical applications.

Our Approach

Explainability of deep learning algorithms is essential to increase the reliability and, thus, enable the deployment of machine learning in safety-critical applications such as autonomous driving. Understanding and explaining the network behavior, providing mathematically sound uncertainty bounds, and increasing the robustness, e.g. against adversarial examples, are major milestones.

Fusion of multiple sensor modalities as well as the capability to distinguish classes without significant labeling effort at high performance and low manual effort are key to enable rich perception.


Our lead application addresses multi-task perception deep networks for video and radar signal processing for automotive video and radar sensors. We tackle challenges from fusing different sensor outputs into a joint environment perception to defining tasks and creating networks unifying detection and semantic segmentations.

Read more

  • Bosch BCAI - Publication - ICLR

    Metzen et al.

    "On Detecting Adversarial Perturbations"
    • Authors: Jan Hendrik Metzen, Tim Genewein, Volker Fischer, and Bastian Bischoff
    • Published in ICLR in 2017
  • Boch BCAI - ICLR

    Fischer et al.

    "Adversarial examples for semantic image segmentation"
    • Authors: Volker Fischer, Mummadi Chaithanya Kumar, Jan Hendrik Metzen, and Thomas Brox
    • Published in Workshop Track ICLR in 2017
  • Bosch BCAI - Publication - ESANN

    Wagner et al.

    "Learning Semantic Prediction using Pretrained Deep Feedforward Networks"
    • Authors: Jörg Wagner, Volker Fischer, Michael Herman, and Sven Behnke
    • Published in ESANN in 2017
  • Bosch BCAI - ICCV

    Metzen et al.

    "Universal Adversarial Perturbations Against Semantic Image Segmentation"
    • Authors: Jan Hendrik Metzen, Mummadi Chaithanya Kumar, Thomas Brox, and Volker Fischer
    • Published in ICCV in 2017
  • Bosch BCAI - Publication - ICIP

    Gondal et al.

    "Weakly-Supervised Localization of Diabetic Retinopathy Lesions in Retinal Fundus Images"
    • Authors: Waleed M. Gondal, Jan M. Köhler, Rene Grzeszick, Gernot A. Fink, and Michael Hirsch
    • Published in ICIP in 2017
  • Bosch BCAI - Publication - ICML


    "Minimum Regret Search for Single- and Multi-Task Optimization"
    • Authors: Jan Hendrik Metzen
    • Published in ICML in 2016
  • Bosch BCAI - Publication - ESANN

    Wagner et al.

    "Multispectral Pedestrian Detection using Deep Fusion Convolutional Neural Networks"
    • Authors: Jörg Wagner, Volker Fischer, Michael Herman, and Sven Behnke
    • Published in ESANN in 2016

Research Application

Bosch BCAI - Universal Adversarial Perturbations Against Semantic Image Segmentation

Universal Adversarial Perturbations Against Semantic Image Segmentation