SupervisedImmuneNet: Training Artificial Immune Networks using a Supervised Learning Approach for Improved Multi-Class Classification

The most common application of artificial immune networks (AINs) is on unsupervised learning tasks. This is due to the fact that AINs are inspired by the adaptive immune system, which consists of a network of antibodies that self-organises to form a memory of external antigens. The self-organising nature of AINs makes them a natural approach for solving problems involving learning and adapting to patterns or structures present in a dataset to form an abstract representation. Training AINs in this fashion means that the dataset need not have class labels because the typical aim of the learning process is not to perform classification. However, there have been attempts to use AINs for classification tasks by considering the resulting clusters of antibodies as representative of the classes present in a dataset. This has also been done when applying AINs to the task of recognising handwritten characters. However, in all the approaches found in the literature, the common method was to leave the task of discovering classes to the AINs. Doing so is contrary to how other models are trained to do classification tasks where data samples are provided along with their class labels to guide the learning process. Therefore, this paper presents a novel supervised learning approach to training AINs for multi-class classification. The proposed approach was tested on the MNIST handwritten digits dataset and achieved a classification accuracy of 99.45%.