Jean-Paul van Oosten - Separability versus Prototypicality in handwritten word retrieval

When using Google to search for a document, you want the top results in the hit list to be most relevant to your search query. The same is true for searching through handwritten documents. In the Monk methodology, the quality of the hit list directly determines the learning rate of the systems. The easier human users can label the word hypothesis, the higher the rate of agglomeration of image instances for a word class.

State-of-the art machine learning methods for ranking and classifying images, however, did not yield satisfactory results. Analysis of the problems with problematic hit lists revealed that discriminatory classifiers may be missing a point. We found out that actually two functions need to be optimised in image retrieval systems: separability and the degree of prototypicality. Our method, ranking images in two stages is convenient for massive scale, continuously (24/7) trainable retrieval engines such as Monk. The new insight is that methods that perform well with regard to separability, do not by definition perform well with regard to prototypicality and vice versa.