Neural network designer
Drag layers, inspect connections, and build a valid architecture.
Architecture inspection
Review layer order, shapes, parameters, and live validation.
Training laboratory
Tune optimisation, regularisation, and validation behaviour.
Training controls
Run, pause, resume, step, stop, or reset simulated learning.
Plotly.js analytics
Inspect learning, complexity, confusion, gradients, and activations.
Prediction tester
Enter features, select samples, and inspect model confidence.
Score, rewards, and feedback
Balance accuracy, speed, simplicity, and generalisation.
Architecture quiz
Answer a contextual question for bonus experience.
Attempt history
Compare architectures, metrics, scores, and completion time.
| Date | Task | Architecture | Parameters | Validation | Score | Mode |
|---|
How to play
Build, validate, train, diagnose, and improve.
1. Select a mission
Choose a task, difficulty, and challenge mode. Read the input, output, target, and resource constraints.
2. Build the network
Add input, hidden, regularisation, and output layers. Keep the output shape compatible with the selected task.
3. Tune training
Select an optimiser, loss, learning rate, batch size, epochs, regularisation, validation split, and early stopping.
4. Read diagnostics
Use loss, score, confusion, probability, complexity, activation, gradient, and generalisation plots to refine the model.
Frequently asked questions
Understand the game’s educational model.
Does this game train a real production model?
No. It uses a deterministic educational simulation that reacts to architecture and training choices.
Why must the output layer match the task?
Binary, multiclass, and regression tasks require different output sizes, activations, and loss functions.
What causes simulated overfitting?
Excessive capacity, weak regularisation, small datasets, long training, and missing augmentation can widen the validation gap.
What causes simulated underfitting?
Too little capacity, poor activations, inadequate epochs, and unsuitable learning rates can prevent useful learning.
How are parameters estimated?
Dense layers use input units multiplied by output units, plus one bias per output unit. Other layers use simplified estimates.
Can progress be saved?
Yes. The game stores architecture, settings, rewards, and history in browser local storage.
Can I upload my own data?
You can load CSV, JSON, or text as fictional feature samples. No server upload occurs.
What is the best architecture?
There is no universal best network. Strong designs match the task while controlling complexity and generalisation.
How do challenge bonuses work?
Bonuses reward compact networks, stable validation, correct outputs, fast training, and efficient parameter use.