AI CLUB PROJECTS / INTERACTIVE LEARNING

Teach a network.
See it learn.

Live Neural Network with Visual Feedback

Start with just two neurons. Build your hidden layers, train on handwritten digits, and discover which design learns best.

01Build your network

Each hidden layer learns patterns from the layer before it.

Up to 6 layers · 1–256 neurons per layerTraining epochsOne epoch is one pass through all 50,000 training digits. Try 1–50; more epochs take longer.

Same challenge for every experiment:
50,000 training digits · 10,000 validation digits
Seeded weights · Fixed data order
Adam optimizer · Batch size 32 · Dropout 20%

What am I changing?

Neurons & hidden layers

Neurons combine input numbers. More neurons give the network more capacity to learn patterns. Extra layers combine those patterns—but bigger is not always better.

Activation functions

ReLU: keeps positive values and turns negative values into zero. A useful starting point.
Sigmoid: squeezes values between 0 and 1.
Tanh: squeezes values between −1 and 1.
These functions help a network learn more than straight-line relationships.

What does accuracy mean?

The percentage of 10,000 validation digits the network labels correctly. Those digits are not used to update its weights. The separate official test set is saved for final judging.

02Your network at a glance

Diagram shows up to 12 dots per layer; labels show the actual sizes.

03Train, test, discover

Ready to train

Ready for your first experiment.

—

Validation accuracy

Your score will appear after training.

—

Training time

—

Change vs. your first experiment

See what it gets right—and wrong

A selected sample: up to 6 correct and 6 incorrect predictions, not a representative accuracy sample.

Train your network to reveal its predictions.

Blue = correct predictionRed = incorrect prediction

Your experiments

Change one thing at a time. Can you explain why your score changed?

Run	Hidden layers / activations	Epochs	Accuracy	Parameters	Time

Your first experiment starts with 2 neurons. Where will you take it?

Try your own handwriting

Draw one digit from 0–9 with your mouse or finger, then ask your trained network to guess.

What the model sees

—

Network's guess

Train a network to enable predictions.

Drawings are cropped, resized to fit 20 × 20 pixels, and centered by their ink inside a 28 × 28 image. Model scores are not a guarantee of correctness. Your handwriting may differ from the training data.