Adding Knowledge Representation & Reasoning to Machine Learning:  
Why and How

Benjamin Grosof, Kyndi

Microsoft Research, Redmond, Washington, USA
July 30, 2018

The domain-independent core of AI consists of two areas:  knowledge representation and reasoning (KRR); and machine learning (ML). Today, research in these two areas tend to be pursued in a loosely coupled manner, by largely separate research communities.  We analyze the fundamental relationships between KRR and ML, including both why and how to tightly combine KRR and ML for the sake of AI systems' overall competence, reusability, and explainability. We focus in large part on recently developed techniques, including:
- Rulelog, a form of KRR that is expressively powerful yet scales well;
- neural networks ML ("deep learning"); and 
- probabilistic logic networks, a form of KRR that bridges deeply to ML.
We discuss in particular how adding highly explainable KRR to ML can provide significantly more trustability to AI systems,
helping to meet the rapidly rising demands, and expectations, of both users and public regulatory authorities.