Neuro-symbolic AI in the field of computer vision
Neuro-Symbolic AI is a relatively new research direction that deals with the integration of purely data-based ML models (sub-symbolic AI) and model-based approaches (symbolic AI). It combines the strengths of these two approaches to address limitations of the most advanced deep learning methods, such as a lack of transparency.
The TTZ Günzburg focuses particularly on Neuro-Symbolic AI architectures in the field of computer vision, including (facial) emotion recognition. In human-human and human-machine interactions emotions play a significant role, underscoring the societal relevance of this research area.
The current challenges in emotion recognition lie in the complexity of the task, particularly with respect to the individuality and ambiguity of emotions, as well as the necessity for precise data annotations (labelling). Previous Research results show that purely data-based AI Models, which rely on deep learning, exhibit limitations in terms of interpretability, verifiability, abstract reasoning, and transferability to other scenarios due to their numerical complexity.
Our goal is to efficiently, reliably, understandably, and verifiably interpret highly complex emotions. To this end, new architectures must be designed, tested, and experimentally validated using specially collected data. This approach aims to improve the reasoning process of such AI systems and more effectively address the complexity of emotion recognition.
