Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Safe Deep Learning for Robots in the Wild
Deep learning is part of more and more solutions in modern robotics. However, it has been shown that neural networks are not as robust to conditions not seen during their training. In this project, we aim for a reliable deployment of deep learning techniques on robots.
Keywords: robotics, deep learning, computer vision, machine learning
Deep learning is part of more and more solutions in modern robotics. However, it has been shown that neural networks are not as robust to conditions not seen during their training. Especially in computer vision, neural networks fail to generalize over different lighting conditions and tend to mishandle difficult scenes by ignoring and/or misclassifying objects with high confidence [1]. In this project, we aim for a reliable deployment of deep learning techniques on robots. Autonomous robots can encounter unknown objects at any time in their natural environment that cannot be modelled during training. We want to investigate the behaviour of neural networks on these situations. Therefore the aim of this thesis is to develop failure detection mechanisms that reliably detect failures. These mechanisms will enable robots to employ promising deep learning techniques in a safe and robust way.
This thesis build up on previous research from e.g. fishyscapes.com or youtu.be/tFflr6k81Hc .
The project develops advanced uncertainty estimation techniques and tests the method in real-world robotic scenarios.
[1] D. Hendrycks and K. Gimpel, “A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks,” in ICLR 2017, 2016.
Deep learning is part of more and more solutions in modern robotics. However, it has been shown that neural networks are not as robust to conditions not seen during their training. Especially in computer vision, neural networks fail to generalize over different lighting conditions and tend to mishandle difficult scenes by ignoring and/or misclassifying objects with high confidence [1]. In this project, we aim for a reliable deployment of deep learning techniques on robots. Autonomous robots can encounter unknown objects at any time in their natural environment that cannot be modelled during training. We want to investigate the behaviour of neural networks on these situations. Therefore the aim of this thesis is to develop failure detection mechanisms that reliably detect failures. These mechanisms will enable robots to employ promising deep learning techniques in a safe and robust way.
This thesis build up on previous research from e.g. fishyscapes.com or youtu.be/tFflr6k81Hc . The project develops advanced uncertainty estimation techniques and tests the method in real-world robotic scenarios.
[1] D. Hendrycks and K. Gimpel, “A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks,” in ICLR 2017, 2016.
- investigation of existing methods that analyze embedding spaces in the network
- extension taking into account more information on e.g. object boundaries, time correlation, class-specific embeddings
- evaluation of the metrics on a real-world robotic scenario
- investigation of existing methods that analyze embedding spaces in the network - extension taking into account more information on e.g. object boundaries, time correlation, class-specific embeddings - evaluation of the metrics on a real-world robotic scenario
- be curious about recent research in deep learning
- strong self-motivation and critical mind
- experience with deep learning frameworks (e.g: Tensorflow, PyTorch)
- students from outside of D-MAVT (particularly, from D-INFK, D-ITET, D-PHYS, and D-MATH) are also highly encouraged to apply
- be curious about recent research in deep learning - strong self-motivation and critical mind - experience with deep learning frameworks (e.g: Tensorflow, PyTorch) - students from outside of D-MAVT (particularly, from D-INFK, D-ITET, D-PHYS, and D-MATH) are also highly encouraged to apply
If you are interested, please send your transcripts and CV to Hermann Blum (blumh@ethz.ch).
If you are interested, please send your transcripts and CV to Hermann Blum (blumh@ethz.ch).