Slide 1MotivationObjectiveArchitectureDetails of ArchitectureDetails of ArchitecturePerformancePerformance Comparison with Traditional Classification MethodsThank youHandwritten Digits Recognition using Multilayer PerceptronYang, LuyuPostal service for sorting mails by the postal code written on the envelopBank system for processing checks by reading the amount of money using computersMotivationDesign artificial neural network for handwritten digits recognitionDevelop proposed network using training samples in MNIST database Achieve good testing resultObjectiveArchitectureAssume each pixel in the digit image is either black or white, which contains 1 bit information. To convert the binary sequence into decimal system {0, 0.1, … ,0.9}, the number of elements should be at least log10(2784), approximately 236.0075. Let the higher precision in MATLAB compensate for the gray edge of stroke. Therefore, 237 neurons are used in the hidden layer to store patterns of training samples. Details of ArchitectureIn the original 784 pixels, even two samples which are very close under Euclidean norm may represent different digits. So quasi nearest neighbor classifier operates at the output of the hidden layer, where the key features are supposed to be more pronounced. Details of ArchitecturePerformance0 2 4 6 8 10 12 14 16 18 2010-410-310-210-1100Training Errorepoch0 2 4 6 8 10 12 14 16 18 2010-210-1100Testing ErrorepochClassifier Testing ErrorMLP + Quasi Nearest Neighbor1.75%K-nearest neighbors using Euclidean norm15.0%Principle component analysis13.3%Radial basis network13.6%Support vector machine using Gaussian kernel11.4%2-layer MLP of 300 units14.7%3-layer MLP of 300+100 units13.05%3-layer MLP of 500+150 units12.95%Performance Comparison with Traditional Classification Methods 1. Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, "Gradient-based learning applied to document recognition," Proceedings of the IEEE, vol. 86, no. 11, pp. 2278-2324, Nov. 1998.THANK
View Full Document