Meet MASKS: A novel Multi-Classifier's verification approach
This work addresses error elimination in classifiers for applications requiring high safety, though it appears incremental as it builds on existing ensemble and verification techniques.
The paper tackles the problem of classifier error reduction by introducing a multi-agent ensemble verification method, achieving an error rate reduction to about one-tenth of individual classifiers on datasets like Fashion-MNIST, MNIST, and Fruit-360.
In this study, a new ensemble approach for classifiers is introduced. A verification method for better error elimination is developed through the integration of multiple classifiers. A multi-agent system comprised of multiple classifiers is designed to verify the satisfaction of the safety property. In order to examine the reasoning concerning the aggregation of the distributed knowledge, a logical model has been proposed. To verify predefined properties, a Multi-Agent Systems' Knowledge-Sharing algorithm (MASKS) has been formulated and developed. As a rigorous evaluation, we applied this model to the Fashion-MNIST, MNIST, and Fruit-360 datasets, where it reduced the error rate to approximately one-tenth of the individual classifiers.