5 steps · 0 done
The digits dataset contains 1,797 handwritten digits scanned at the University of California, Irvine, each reduced to an 8×8 grid of pixel intensities from 0 to 16. It's a small cousin of MNIST, small enough to train on in your browser in seconds.
You'll write a two-layer network (64 inputs → 64 hidden ReLU units → 10 outputs), derive and check its gradients, and train it with mini-batch gradient descent. No frameworks: every number is yours.
The project uses everything in this module. Solve these first:
- 4.1 Theory test: Neurons and networks
- 4.1 Practical: A forward pass
- 4.2 Theory test: Backpropagation
- 4.2 Practical: Backward through a layer
- 4.3 Theory test: Training deep networks
- 4.3 Practical: Adam and He initialisation
what you'll build
- Look at the data
- Initialise and run forward
- Backpropagation, checked
- Train with mini-batches
- Study the mistakes