P. Hitzler, Md Kamruzzaman Sarker
An integrated framework combining neural networks and symbolic reasoning achieves both data efficiency and interpretability.
Existing neural networks are difficult to interpret due to data dependency and black-box nature, while symbolic AI lacks flexibility. A method to integrate the strengths of both paradigms is needed.
Design a bridge architecture connecting low-level pattern recognition of neural networks with high-level logical reasoning of symbolic systems. Specifically, introduce a neural-symbolic interface that converts neural network outputs into symbolic representations, and conversely, utilize symbolic rules in neural network training.
The proposed framework achieves high performance with less data compared to existing neural networks and provides interpretability that explains the reasoning process. Additionally, it enables knowledge injection through symbolic rules, making it effective for domain-specific problems.