Neuro-Symbolic AI
Historically, there have been two broad classifications in the field of artificial intelligence: Symbolic and Connectionist AI.
Symbolic AI is the epitome of conceptual and logical clarity and is based on the formalisation of knowledge into facts, rules and other symbolic structures. A symbol is a pattern that represents other things. The target can be an object, another symbol or a relationship. Imagine a large library where each book is a rule and each chapter is a path to logical deduction - symbolic AI, analogous to the thought processes of human reasoning.
Connectionist AI, inspired by the structure of the human brain, involves artificial neural networks and machine learning algorithms that represent knowledge in terms of distributed activation patterns across a network of units. This area has become increasingly important in recent years due to growing data volumes and computing power and is characterized by robustness.
Connectionist models are well suited for “best-match” problems and are more flexible than symbolic representations as they can handle incomplete, approximate and inconsistent information as well as exceptional situations. . However, they require a large amount of training data to capture simple concepts that humans can grasp with few or single examples.
| Symbolic AI | Connectionist AI |
|---|---|
| Explicit knowledge | Implicit knowledge |
| Explainable | Black Box Nature |
| Leverage existing knowledge | Pattern recognition |
| Reasoning capabilities (e.g. logic, rules) | Fault tolerance |
| Knowledge Engineering and Acquisition Bottleneck | Huge amount of training data required |
| Continual Learning | Catastrophic Forgetting |
Recently, it has been increasingly recognized that the two approaches are not mutually exclusive. There are more and more attempts to combine both methods in order to utilize their respective strengths - an approach that has become established as “neuro-symbolic”. In the field of image recognition, for example, connectionist models driven by deep neural networks demonstrate their strength in recognising subtle patterns in pixel data, similar to the way our brain recognises faces in a crowd. They also prove their strength in medical imaging. In expert systems for medical diagnosis, on the other hand, symbolism shines through the methodical application of predefined rules in the diagnosis of diseases, imitating the logical sequence of a doctor's thought process. Combined, for example, the classification of images using connectionist models can serve as a basis for further processing by symbolic AI. Another example of the application of neuro-symbolic paradigms is the interaction of large language models (LLMs) and knowledge graphs. LLMs excel in handling natural language and common-sense reasoning, while knowledge graphs are used to process explicit domain knowledge.