Integrating Deep Learning Based Perception with Probabilistic Logic via Frequent Pattern Mining

The bridging of the gap between 1) subsymbolic pattern recognition and learning algorithms and 2) symbolic reasoning algorithms, has been a major issue for AI since the early days of the field. One class of approaches involves integrating subsymbolic and symbolic systems, but this raises the question of how to effectively translate between the very different languages involved. In the approach described here, a frequent subtree mining algorithm is used to identify recurrent patterns in the state of a hierarchical deep learning system (DeSTIN) that is exposed to visual stimuli. The relationships between state-subtrees and percepts are then input to a probabilistic logic system (OpenCog’s Probabilistic Logic Networks), which conducts uncertain inferences using them as axioms. The core conceptual idea is to use patterns in the states inferred by a perceptual hierarchy, as inputs to an uncertain logic system. Simple illustrative examples are presented based on the presentation of images of typed letters to DeSTIN. This work forms a component of a larger project to integrate perceptual, motoric and cognitive processing within the integrative OpenCog cognitive architecture.