COMP 551 Project 3 - Multilabel classification of handwritten digit images
This project implements a deep neural network model for a multi-label classification problem on a modified MNIST dataset. The dataset was pre-processed, and the model and hyperparameters (optimizer, learning rate, batch size, number of epochs) were tuned through manual testing and exhaustive grid search to improve accuracy. The combination of hyperparameters resulting in highest validation accuracy was picked, resulting in an accuracy of 99.785 on the public test set on kaggle.
Done with Shayan Sheikh and Matthew Kourlas for COMP 551: Machine Learning at McGill University.