Neuro-Symbolic AI is a hybrid approach to artificial intelligence that combines the strengths of both symbolic and connectionist models. Symbolic AI is based on the manipulation of abstract symbols and logical reasoning, while connectionist AI is based on the simulation of neural networks and learning from data. Neuro-Symbolic AI seeks to overcome the limitations of each approach by integrating them into a single system that can reason about both symbolic and subsymbolic representations. This approach has shown promise in areas such as natural language processing, robotics, and combined model-driven and data-driven control, where the combination of symbolic reasoning with neural network-based perception and learning can lead to more robust and interpretable AI systems.
GPT-3/4 and BERT are well-known examples of Neuro-Symbolic applications: Words or subwords are both the input and output symbols of these large language models. Internally, subsymbolic computation happens. Another well-known example is AlphaGo, where the symbolic techniques call neural techniques, employing Monte Carlo Tree search to calculate on board states as symbolic information.
Within IDLab, we work on multiple Neuro-Symbolic approaches, including large language models, transformers and combining graph-based reasoning and modelling.
