Fashion MNIST Image Classifier

An image classifier for the Fashion MNIST dataset, 28x28 grayscale images across ten clothing categories, built as a first end-to-end deep learning exercise.

Model

Pixel values are normalised, then a Flatten layer converts each image into a one-dimensional vector, a hidden layer of 128 ReLU neurons extracts features, and a softmax output layer of 10 neurons produces the class distribution. Training uses the Adam optimiser with sparse categorical cross-entropy over five epochs.

Evaluation

The model reaches 86.8% accuracy on the test set. Beyond that number, predictions are visualised alongside the images with both actual and predicted labels, because looking at the specific mistakes is more informative than the aggregate: the confusions concentrate in the categories that genuinely look alike at 28x28, like shirts against coats and pullovers.