ATHENA: Advanced learning through efficient transparent model harnessing with explainable data

Context

In cyber-physical systems, such as highly automated vehicles, autonomous robots and the like, the use of neural networks is particularly interesting due to their immense performance in dealing with high-dimensional data. Even with modern architectures, these models require considerable effort to provide the necessary data points for training. Despite efficient training strategies, the data must be laboriously labeled, an expensive and tedious task. In addition, an immense amount of energy has to be expended for training, which cannot be expedient from a sustainability perspective. xAI (explainable AI) methods, however, allow both the post-hoc explanation of why a model has made a certain decision and the development of self-explanatory models. This allows targeted investigations to be carried out and important data points to be extracted. The aim of the work is to select training data that has particular added value for the model and thus for retraining. To this end, various xAI methods will be researched, applied and evaluated against each other.

Targets

• Research on explainable AI methods (xAI), in particular on intrinsically explainable neural networks
• Selection of methods
• Application of the selected methods to training data
• Implementation of an end-to-end workflow (pipeline with found data points) for efficient retraining
• Evaluation and benchmarking of the methods and results

Requirements

• An independent and structured way of working is particularly important
• Programming skills in Python
• Interest in xAI
• Knowledge of game theory, linear algebra, etc. is desirable
• Knowledge of PyTorch or Tensorflow