Eliciting Domain Knowledge in Handwritten Digit Recognition

Pattern recognition methods for complex structured objects such as handwritten characters often have to deal with vast search spaces. Developed techniques, despite significant advancement in the last decade, still face some performance barriers. We believe that additional knowledge about the structure of patterns, elicited from humans perceptions, will help improve the recognition’s performance, especially when it comes to classify irregular, outlier cases. We propose a framework for the transfer of such knowledge from human experts and show how to incorporate it into the learning process of a recognition system using methods based on rough mereology. We also demonstrate how this knowledge acquisition can be conducted in an interactive manner, with a large dataset of handwritten digits as an example.