HESIP: A Hybrid System for Explaining Sub-symbolic Predictions

Abstract

Machine learning models such as neural networks have been successfully used in many application domains such as mission critical systems, digital health and autonomous vehicles. It is important to understand why particular predictions are made by a sub-symbolic machine learning (ML) model, because humans use these predictions in their decision making process. In this paper, we introduce HESIP, a hybrid system that combines symbolic and sub-symbolic representations to explain a prediction in natural language for an image prediction task. A sub-symbolic ML model makes a prediction for an image, and based on this predicted image, the system selects sample images from the dataset. Afterwards, a symbolic ML model learns probabilistic rules using the representation of positive and negative sample image instances where the decision about a positive or negative image instance comes from the sub-symbolic ML model. The prediction of an image can then be explained in natural language using the learned rules. Our evaluation shows that the probabilistic rules can be learned with high accuracy.