Neural Networks by Hand in Excel

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Build one artificial neuron that decides which Lumora Coffee Club subscribers will renew, measure its accuracy and find its decision boundary.
• Build the four classic activation functions side by side and the sigmoid's slope, which backpropagation needs later.
• Push six Lumora customers through a small network with two inputs, two hidden neurons and one output, and read off renewal probabilities.
• Score eight renewal predictions with mean squared error and with cross-entropy, and find the prediction that hurts the network most.
• Run one full backpropagation step for one Lumora customer: error signals, all nine gradients via the chain rule, the weight update, and you watch the loss fall.
• Train the same one-weight model with three learning rates for five steps, chart the loss curves and see crawling, converging and exploding training.
• Turn the network's raw scores for six Lumora support tickets into probabilities over three classes, pick the predicted class and score it with categorical cross-entropy.
• Track Lumora's churn network over ten epochs, measure the gap between training and test error, pick the best epoch with INDEX/MATCH and chart both curves.