Bidirectional Negative Correlation Learning

Abstract

Negative correlation learning method is to create different individual learners for building a committee machine. In the original version of negative correlation learning, the learning target on a give data point was set to be the same for all the individual learners in the committee. The same learning target could lead the individual learners to become similar to each other if the learning process would be conducted for long. In order to create more different and cooperative individual learners for a committee machine, different learning targets should be set on each learning data for different individual learners in negative correlation learning. In this paper, negative correlation learning with two different learning targets was implemented. On learning each training data, the individual learners could go to the two different learning directions so that there would be little chance for them to become similar even if a long learning process would be performed. Experimental results would show how the two different learning targets would allow the individual learners to become both weak and different in negative correlation learning.