Project 4: Convolutional Neural Networks

Taige Hu

This project, we used a convolutional neural network to classify images.

Part 1: Image Classification

We looked at the fashion MNIST database and I used a neural net structure to try to classify the test sets. I used two convolutional layers with maxpool and ReLU followed by two fully connected layers. I used Adam to optimize the training parameters.

Here are the training and validation accuracies through the epochs of training the net.

I was able to achieve a test accuracy and validation accuracy of 90% after 10 epochs.

The accuracies per class are listed below.

The hardest classes to train were shirt, pullover and tshirt because they are so similar in appearance and have very specific differences that may be hard to quantify. They also span a range of different forms within the class that might overlap among other classes.

Here are some examples of successes and failures from each class.