Explaining the predictions of machine learning models
Systems that use some form of Artificial Intelligence (AI) play an increasing role for making decisions in our daily life. Machine learning techniques; especially, sub-symbolic techniques are often used in these AI systems to find solutions for prediction tasks. While using these AI systems, people often receive answers that they do not understand and they have problems to figure out why a particular decision has been made by the system. This is because these AI systems often employ backbox techniques that are not easily interpretable.
The literature shows that relation information plays an important role and can be determined from the location information of objects in an image and this helps to better understand predictions. These explanation systems explain predictions using the symbolic rules that are learned employing the information represented in an informal way. Consequently, there is no way a domain independent approach can be used in these systems to generate the explanations. We have not found an explanation system that allows human users to modify the explanations for improving the learning process.
To overcome these issues, we propose a novel hybrid explanation system called HESIP for sub-symbolic image predictions. HESIP combines a sub-symbolic and a symbolic component that communicate via an interface layer to learn the explanatory rules. This thesis makes the following contributions: in the interface layer, we use an approach for extracting image information that allows us to easily apply it to a new problem; in the symbolic component, the image information including the relation information between the objects in the image is formally represented employing an ontology and the explanatory rules are learned using this information; we use a bi-directional grammar for generating natural language explanations that are human-understandable and machine-processable, and the explanation generation process does not depend on the problem domain; and finally, we support a humanin- the-loop process for modifying the generated explanations.
HESIP explains image predictions where an explanation for a predicted image represents the object information together with the relation information. The subsymbolic component makes a prediction for an image. The symbolic component then learns probabilistic symbolic rules from the sample positive and negative image instances. These instances are selected based on the predicted image whereas the sub-symbolic component provides the decisions about these instances. An ontology is used to represent the information of the sample images in the symbolic component for rule learning. HESIP generates explanations in a controlled natural language using the learned probabilistic rules employing a bi-directional logic grammar. HESIP allows a human-in-the-loop to modify an explanation if the user finds that the explanation is not correct. During this modification process, the user can modify the information of the controlled natural language explanation. Afterwards, HESIP learns a new explanatory rule in the symbolic component and the newly learned rule is used to generate an explanation that better explains the image prediction. Our evaluation results show that HESIP can generate explanations with high accuracy and the explanation modification process can effectively learn better explanations.
