Advanced Testing of Deep Learning Models

This course will give you an opportunity to implement and learn state-of-the-art testing strategies for AI-based software systems. The wide range of AI software systems is applied in safety-relevant areas such as autonomous driving and is, therefore, rapidly becoming part of our daily lives. This created the need for researchers to develop testing strategies for various kinds of machine-learned systems such as Deep Neural Networks (DNNs). In this practical course, you will implement testing algorithms for Deep Neural Networks inspired by advanced software testing strategies like coverage-guided fuzzing and latent space-aware testing. Testing strategies considering the latent space of DNNs are becoming popular as latent space plays a crucial part in decision-making. Also, the modeling and testing of latent space could help us understand the behavior of DNNs in corner cases. The students will work in teams in order to develop testing frameworks for deep neural networks used in 2D/3D object detection and classification. These DNNs would typically be evaluated on benchmark datasets like A2D2, KITTI, BDD100K, GTSRB, SVHN etc. depending on the model's application domain. The nature of the task allows you to work remotely from your comfort place. Supervisor: Simon Speth, Vivek Vekariya

Requirements

1. Python, Linux

2. Deep Learning Frameworks (PyTorch, Keras, TensorFlow etc.)

3. Classification and 2D Object Detection Networks (such as CNNs, Spatial Transformer Networks, SSD300, Faster-RCNN etc.)

Good-to-have

1. Knowledge of advanced software testing methods like latent space-aware testing & fuzz testing

2. Deploying code on GPU Servers

3. Adversarial attacks on object detection networks