Border Pairs Method (BPM) is a new constructive method for supervised learning of multilayer perceptron (MLP), which calculates, values of weights and biases directly from the geometry of learning patterns.
To determine BPM’s capabilities, we compared it with three other supervised machine learning methods: Backpropagation , SVM and Decision Trees. The comparison were made on six databases: XOR, Triangle, Iris, Pen-Based Recognition of Handwritten Digits, Online Pen-Based Recognition of Handwritten Digits and synthetically generated noisy data. Border Pairs Method found near minimal MLP architecture in all described cases. For classification of the Iris Setosa only two border pairs (only four patterns out of 150) were enough for learning the whole data set correctly. In the classification of ‘Pen-Based Recognition of Handwritten Digits’ dataset only 200 learning patterns were used for learning. The BPM correctly identified more than 95% from 3498 handwritten digits, which did not participate in the learning proces. We are particularly satisfied with the results of online learning. The adaptability of the BPM to new incoming patterns is so high, that it easily follows and sometimes even exceeds the offline version of the algorithm. Even with a noisy data, some very good results were obtained. RMSE error of the BPM was significantly lower than that of Backpropagation, especially for the noise lower than 10%.
The initial results of the research are good starting point for further research work. Some possible directions are:
· Noise reduction using border pairs.
· Algorithm improvement - better integration of border pairs.
Link to the scientific paper
Link to the Book
Bojan PLOJ, PhD