Statistical Methods for Machine Learning experimental project
This repository presents an implementation of the Kernel Perceptron algorithm built from scratch, to perform multiclass classification on the MNIST database of handwritten digits. We run the algorithm for a different number of epochs and degrees of the polynomial kernel, to obtain two types of "One-versus-All" predictors from the ensemble: the average of the predictors and the predictor which achieves the smallest error. The performance of the predictors is evaluated on the test set, using Accuracy, Precision and Recall metrics.
The MNIST dataset can be downloaded from Kaggle.