Module 04 Activity

Scenario

You have been asked whether a neural network is worth deploying for a structured-data problem. Answer with evidence rather than preference.

Task

1. Show that two stacked linear layers with no activation collapse to one.
2. Fit the circle task at hidden widths 0, 4, 16 and 64. Record train and validation accuracy for each.
3. Build the three head types and confirm their output shapes and that multiclass probabilities sum to 1.
4. Scale badly-scaled features and record the accuracy difference.
5. Fit a majority-class baseline and logistic regression, then the network. Report all three.

Deliverable

A comparison table with the three scores, the width where validation stopped improving, and a verdict.

Check your work

• The stacked pair and the equivalent single layer produce identical output.
• Width 0 cannot represent a circle. Validation flattens well before training does.
• Multiclass probabilities sum to 1.0 per row.
• Scaled features beat raw features with identical architecture and step count.

The verdict to write

If the network only matches logistic regression, say so. That is a result, and it saves deploying complexity that bought nothing.

Mark complete