Verification of Deep Learning Algorithms for Autonomy Related Tasks

Project started September 2019.

Deep learning has rapidly become the backbone of a variety of algorithms that are used for perception within autonomous systems [1]. Despite their superior performance across a variety of different perception related tasks that are crucial to autonomous automotive applications across a variety of benchmarks, techniques to validate their accurate performance remain lacking. Though incredibly successful, these discriminative neural network-based approaches are surprisingly susceptibility to subtle perturbations to their inputs, which can lead to incredibly poor performance [2]. For example, given the task of classifying ground terrain from images or predicting the future location of a pedestrian given their prior locations, state-of-the-art deep learning techniques can be completely fooled by perturbations to the input data that are often completely imperceptible to a human. Typically, the adversarial examples that fool these discriminative neural networks are constructed by generative neural networks that are trained explicitly to create such examples [3].

This project will develop methods for testing and validation of deep learning algorithms used for autonomy. More specifically, this project will develop novel optimization-based techniques to verify that a given discriminative neural network is robust to perturbations that are bounded up to some user specified threshold. In contrast to existing approaches, the novelty of this technique will be in its ability to obtain an adversarial example that is within the user-specified threshold, if one exists, or a certificate of robustness if such an example does not exist.

References:

1. LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. “Deep learning.” Nature 521.7553 (2015): 436.
2. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., and Fergus, R. (2013). Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199.
3. Radford, Alec, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015).