Computer Science 4033/5033: Machine Learning
Classify handwritten digits using machine learning techniques
Yan Liang, Yunzhi Wang and Delong Zhao
Fall 2015
Project scope For our machine learning project, we propose to build several machine learning classifiers that recognize handwritten digits. Handwritten digit recognition is a classic problem in machine learning studies for many years. We plan to do several experiments using different machine learning algorithms and compare the pattern recognition performance. We hope to create a classifier that has same or better categorization accuracy than record performance from previous studies. Yan will focus on neural network, Delong will focus on the random forests methods, and Yunzhi will focus on SVMs and KNNs. We will also develop a final novel classifier that combines the best models from our different experiments. We hypothesize that the final classifier will archive a categorization accuracy of 0.99. This indicates that the classifier correctly classified all the handwritten digits but 1% of the images.
The goal of handwritten digit recognition is to determine what digit is from an image of a single handwritten digit. It can be used to test pattern recognition theories and machine learning algorithms. Preprocessed standard handwritten digit image database has been developed to compare different digit recognizers. In our semester project, we will use modified National Institute of Standards and Technology (MNIST) handwritten digit images dataset from kaggle digit recognizer project. The Kaggle MNIST dataset is freely available and collected 28,000 training images and 42,000 test images. Each image is a preprocessed single black and white digit image with 28 x 28 pixels. Each pixel is an integer value range from 0 to 255 which represent the brightness of the pixel, the higher value meaning darker. Each image also has a label which is the correct digit for the handwritten image. For each input handwritten image, our model will output which digit we predict and evaluate with the correct label. We will use 28,000 training images to train our machine learning model and use 42,000 test images to test the performance. Then we will calculate the percentage of the test images that are correctly classified and compare the performance of different machine learning algorithms.