Fully Connected Neural Network and Convolutional Neural Network for MNIST

As part of the course MAC5921 - Deep Learning of Instituto de Matemática e Estatística - USP, it was proposed a programming exercise with the intention of promoting greater familiarization with the usage of PyTorch library to train and test Fully Connected Neural Networks and CNNs. For this, it was proposed the usage of MNIST, a dataset that contains images of handwritten digits, divided in 60.000 images for training and 10.000 for testing. For this programming exercise, the following tasks were proposed and accomplished in the code provided:

• Implementation of both NN and CNN (not very large), both with similar number of trainable parameters (NN: XXX | CNN: XXX);
• Inclusion of transformations such as salt-and-pepper noise in the input images alongside an vizualisation of its impact on the results of both models;
• Analysis of the impact of unbalanced classes on models performance;
• Results vizualizations through Loss Curves, Confusion Matrices and t-SNE plots;
• CNN's focus on images vizualization through GradCAM.