NN-1 · Core Activity · 35–45 minutes

Make it fail, then make it learn

Your job is to be a scientist: predict what will happen, run the experiment, then explain what you saw. You will deliberately make a neural network fail before you make it succeed.

Golden rule: a wrong prediction with an honest, well-reasoned explanation is useful evidence. Do not change a prediction after seeing the result.

Learning focus
Non-linearity, hidden capacity, loss, generalization

What to submit
Your predictions, observations, explanations, and reflection

Response storage
Autosaves locally in this browser only

Ready. Responses save locally as you type.

The controls you will use

Control	Use in this activity
Dataset	Two moons · Concentric circles · XOR · Two blobs
Activation	tanh · ReLU · sigmoid · linear
Hidden 1 / Hidden 2	0–12 neurons; 0 means no hidden layer
Learning rate / optimizer	Keep defaults unless a run says otherwise
Train / pause · Reset weights	Run and repeat experiments
Show test loss/points	Compare training behavior with held-out data

Part A · Five experiments

For every run: Predict first → Run it → Explain what happened.

Run 1 · Can a straight line bend?

Give the network a curvy problem while switching off its ability to create a nonlinear transformation.

Set up: Dataset = Two moons · Activation = linear · Hidden 1 = 6 · Hidden 2 = 4 → Train.

Predict: Will this network separate the two moons? Why?

Observe: What happens to the decision boundary and loss?

Explain: Why can a stack of linear layers still fail to create a curved boundary?

Run 2 · No hidden layer

Set up: Dataset = Concentric circles · Activation = ReLU · Hidden 1 = 0 · Hidden 2 = 0 → Reset weights → Train.

Predict: Can this network separate the inner dots from the outer ring? Why?

Observe: What does the loss do?

Explain: What limitation is visible in the boundary?

Run 3 · Add capacity

Set up: Keep Concentric circles and ReLU · Hidden 1 = 6 · Hidden 2 = 4 → Reset weights → Train.

Predict: What shape do you expect the boundary to become?

Observe: What changes compared with Run 2?

Explain: What did the hidden layers make possible?

Run 4 · Watch it learn

Set up: Dataset = XOR · Activation = ReLU · Hidden 1 = 6 · Hidden 2 = 4 → Reset weights → Train and watch the whole process.

Predict: What should happen to loss if training is working?

Describe the loss curve over the epochs.

Describe what the decision boundary does as loss changes.

Run 5 · Train versus test

Increase capacity, then ask whether the model learned a reusable pattern or mostly fitted the training examples.

Set up: Dataset = XOR · Activation = ReLU · Hidden 1 = 10–12 · Hidden 2 = 10–12 · turn on Show test loss/points → Reset weights → Train.

Predict: Will training and test loss stay close together? Why?

Observe: What happens to the training and test curves?

Explain: What would a persistent train/test gap mean?

Part B · Use the explanation on the applet page

1. Why does a chain of linear transformations collapse to another linear transformation?

2. In your own words, what happens during one training epoch?

3. Find one common failure mode you did not deliberately test above. What setting causes it?

Part C · Reflection and transfer

1. Traceability: Pick one prediction that was wrong. What did you predict, what happened, and what did it teach you?

2. Counterfactual: What is the most important thing a hidden layer with a nonlinear activation lets the model do that a single linear neuron cannot? Use evidence from a run.

3. Connect forward: Based on what you observed here, predict one way a neural network might learn useful structure from an image.

How this is assessed

Criterion	What strong work shows
Predictions made first	A prediction exists before each run, even when it turns out wrong.
Observations	Specific description of what the boundary and loss actually did.
Explanations	Connections to non-linearity, hidden layers, training, and generalization.
Reflection	Reasoning from evidence rather than restating a definition.

Local response storage

Your typed responses are saved only in this browser using local storage so an accidental refresh does not erase your work. Nothing on this activity page submits answers to AI Playgrounds. Use Clear responses to remove the saved copy from this browser.