Game configuration
Choose how you want to train
Scenario 1
Press Start to begin
You will receive a task, output setup, constraints, and several loss-function choices.
Select the best loss function
Keys 1–4
Start a round to receive immediate explanations, compatibility checks, gradient notes, and implementation guidance.
Canvas simulation
Prediction and penalty view
Round progress
0 of 100 XP
0
Correct0
Hints0
Best streakPlotly analytics
Loss curves and gradient behaviour
Interactive laboratory
Manipulate targets and predictions
Live penalty comparison
Sample table
| Sample | Target | Prediction | Error | Loss | Gradient |
|---|
Custom challenge builder
Create your own loss-selection scenario
Configure the fields, then build a playable custom scenario.
Performance statistics
Your mastery profile
Recent attempts
| Time | Scenario | Selected | Result | Points |
|---|
Rewards
Badges and unlocks
Saved progress
Local storage controls
Progress remains in this browser until you clear it.
Loss function reference
Quick selection guide
| Loss | Best for | Strength | Watch out for |
|---|
Common compatibility rules
- Binary cross-entropy usually pairs with sigmoid probabilities.
- Categorical cross-entropy usually pairs with softmax outputs.
- Sparse categorical cross-entropy expects integer class identifiers.
- Huber and log-cosh reduce extreme outlier influence.
- Dice and Tversky losses focus on segmentation overlap.
- Contrastive and triplet losses learn distances between embeddings.
- Detection systems often combine classification and localisation losses.