Project idea · ungraded

Handwritten Digit Recogniser

Train a neural network that recognises handwritten digits, then try it on digits you draw yourself. Start with the small 8×8 digits that ship with scikit-learn, then move to MNIST's 28×28 images.

You will use

Milestones

Build the smallest working version first. Ticks are saved in this browser only.

01Load the 8×8 digits with sklearn.datasets.load_digits, scale pixels to [0, 1] and split train/validation/test.
02Train an MLP with the training loop from the course and plot train and validation loss per epoch.
03Report test accuracy and a confusion matrix; look at the ten worst mistakes.
04Switch to MNIST through torchvision and train the same model on 28×28 images.
05Write a predict(image) function and test it on digits you draw yourself (scaled to the same size).

Stretch goals

Add dropout and compare validation curves; save the best model with its state_dict.
Replace the MLP with a small CNN (see the Computer Vision topic).
Make a tiny web page or FastAPI endpoint where you draw a digit and get a prediction.

Data

1,797 8×8 digits bundled with scikit-learn (no download): sklearn.datasets.load_digits.
MNIST, 70,000 28×28 digits, downloaded by torchvision.datasets.MNIST.

Run it yourself

pip install torch torchvision scikit-learn matplotlib
python train_digits.py

A CPU is enough: an MLP on MNIST trains in a few minutes.