Builder project·2-3 hours

Neural Network Debugging Project

Repair a broken training loop using shape checks, gradient evidence, a tiny-batch test, and learning curves.

Scenario

A binary MLP remains at chance accuracy. You are given a tiny deterministic case and must identify whether the defect is in the forward pass, loss, gradients, or update.

You will demonstrate

Use invariants to localize training defects.
Verify gradients independently.
Separate optimization health from generalization.

Project evidence

Complete each deliverable.

Deliverable 1

Assert interfaces

Write expected tensor shapes for input, logits, labels, and gradients.

Evidence: Passing shape and finite-value assertions.

Deliverable 2

Check a gradient

Compare the analytical scalar gradient with a centered finite difference.

Evidence: Relative error below the stated tolerance.

Deliverable 3

Overfit a tiny batch

Show that loss can decrease substantially on a deterministic subset.

Evidence: Loss trace and final predictions.

Deliverable 4

Write the diagnosis

Name the defect, evidence, correction, and a regression test.

Evidence: A falsifiable root-cause note.

Runnable Python lab

Repair a broken scalar update

Implement the gradient and update needed to fit a one-weight model.

Test output

Run the code when you are ready. The first load downloads the in-browser Python runtime.

Code runs only in a time-limited browser worker. It is not sent to the IntuitiveAI server. Do not paste secrets.

Expected result

w: approximately 3.0
loss: near 0
ALL TESTS PASSED

Project defense

What should be checked first when a tiny MLP cannot overfit eight deterministic examples?

Assessment rubric

Assertions localize failures.
Gradient parity uses a stated tolerance.
The tiny-batch test is deterministic.
The fix includes a regression check.

Useful references

A Recipe for Training Neural NetworksPractical debugging and validation workflow for neural-network training.

Project completion gate

Complete all deliverables, pass the automated code checks, and defend the key decision.

Deliverables 0/4Code pendingDefense pending