Towards autonomous machine reasoning: Multi-stage classification system with intermediate learning

Abstract

This paper describes a new concept of multi-stage classification with intermediate learning (MSIL), and validates a simple two-stage version of the MSIL on nine popular test datasets. The first stage performs classical learning and inference based on features calculated directly from the data. The second stage learns and infers the final diagnosis using diagnostic labels generated at the first stage. Since both stages are trained independently, the learning results of the second stage do not alter the learning results accomplished at the first stage. This important property enables the generation of more complex, multi-channel and/or multi-level machine reasoning systems consisting of algebraically connected basic two-stage units. Classification tests showed that in almost all tested cases, the accuracy achieved at the first stage was further improved by the second stage of classification. This means that primary learning from the data can be improved by secondary learning from mistakes made when classifying the data parameters.