Neurosymbolic Artificial Intelligence: Theory, Applications, and Challenges

Connectionist models such as deep neural networks have shown great power in learning with big data, achieving state-of-the-art performance in many narrow problems. Recently, much attention has been focused on improving their reasoning capability to solve intellectual problems logically in a larger context. For this challenge, neurosymbolic computing has been emerging as a promising approach where symbol manipulation can be integrated to compensate for the weaknesses of deep neural networks such as opacity and data hungriness. Neurosymbolic offers a systematic amalgamation between learning and reasoning and is considered a potential pathway toward Artificial General Intelligence. In the first part of this talk, I will present the theory behind neurosymbolic computing and the methods to represent symbolic knowledge in neural networks. The second part of the talk will showcase the applications of neurosymbolic and discuss some interesting research challenges in this topic.