Fashion MNIST Classification Using Convolutional Neural Networks

UC Irvine, US

This project focused on developing a Convolutional Neural Network (CNN) to classify images from the Fashion MNIST dataset, which consists of grayscale images of various clothing items. The goal was to build an accurate deep-learning model that could effectively classify fashion items while balancing model complexity and computational efficiency.

The CNN architecture was designed with convolutional, pooling, batch normalization, and dropout layers to extract features and reduce overfitting. The model was trained using TensorFlow and Keras, optimizing it with the Adam optimizer and categorical cross-entropy loss function. It achieved a test accuracy of 91.62%, demonstrating strong generalization. Various visualization techniques, including confusion matrices and test image plots, were used to analyze the model's strengths and misclassifications.

While the model performed well, it faced minor limitations due to the lack of data augmentation and the constraints of a fixed architecture. The project highlights the effectiveness of deep learning in image recognition and suggests future improvements such as more complex architectures or pre-trained models to further enhance performance.