A New Physics-Inspired Discriminative Classifier

Concepts and laws of physics have been a valuable source of inspiration for engineers to overcome human challenges and problems. Classification is an important example of such problems that plays a major role in various fields of engineering sciences. It is shown that discriminative classifiers tend to outperform their generative counterparts, especially in the presence of sufficient labeled training data. In this paper, we present a new physics-inspired discriminative classification method using minimum potential line. To do this, we first consider two groups of fixed point charges (as two classes of data) and a movable classifier line between them. Then, we find a stable position for the classifier line by minimizing the total potential integral on the classifier line due to the two groups of point charges. Surprisingly, it will be shown that the obtained classifier is actually an uncertainty-based classifier that minimizes the total uncertainty of the classifier line. The effectiveness of the proposed method is validated by some experiments on both synthetic and real datasets. First, two synthetic datasets are constructed to visually demonstrate the efficiency of the proposed method. Then, some real-world benchmark datasets are selected (from the well-known UCI machine learning repository) to compare the classification performance of the proposed method with three well-known classification methods.