Machine Learning for Konwledge Injection in Neural Networks

Despite the substantial achievements of artificial neural networks (ANNs) driven by high generalization capabilities, flexibility, and robustness, the training duration and performance still highly depend on several factors, such as the network architecture, but most importantly on the available training data. However, in many real-world applications, e.g., in safety-critical systems, there are various issues regarding data collection and generation. In these scenarios, the capabilities of exclusively data-oriented approaches to train an ANN are limited. To tackle these domain-specific challenges, the concept of integrating or injecting existing domain knowledge into the generation process of ANNs becomes increasingly attractive in research and practice. However, the initialization of ANNs is mostly overlooked in this paradigm and remains an important research question. In our work, we propose a machine learning framework enabling an ANN to perform a semantic mapping from a well-defined, symbolic representation of domain knowledge to weights and biases of an ANN in a specified architecture.