Mission and challenge objectives
Choose a controlled experiment, predict the result, then evaluate the measured trade-offs.
Configure the dataset, model, and training process
Keep most settings unchanged when comparing batch sizes fairly.
Game and challenge
Dataset controls
Model controls
Batch settings
Optimisation controls
Training controls
Make performance predictions
Predictions are scored when the experiment finishes.
Canvas batch-processing simulation
Samples enter batches, gradients update the model, and memory pressure changes with batch size.
Live curves and experiment comparisons
Zoom, pan, hover, and download each interactive graph.
Rank, pin, duplicate, and compare completed runs
Results are stored in this browser and remain available after reloading.
| Rank | Experiment | Batch | Optimiser | Rate | Epochs | Time | Train loss | Validation loss | Accuracy | Stability | Memory | Gap | Score | Actions |
|---|
Automatic findings and learning feedback
The explanation updates after every completed run.
Small batches
They update weights frequently. Their gradients are noisy. Noise can sometimes escape weak solutions.
Large batches
They process samples efficiently. They need more memory. Their gradients are usually smoother.
Fair comparisons
Keep data, model, seed, and optimiser fixed. Change one setting. Compare several repeated runs.
Learning-rate interaction
Larger batches may tolerate larger rates. Excessive rates still diverge. Scaling rules are starting points.
Convergence
Fast epochs do not guarantee fewer epochs. Count updates too. Check validation performance before deciding.
Generalisation
Smooth training can still overfit. Compare validation loss. Prefer balanced performance over one metric.
Achievements
Complete varied experiments to unlock learning milestones.
How to use the game
- Select a challenge and configure one batch size.
- Make predictions before starting the training simulation.
- Observe sample batches, gradient movement, and memory blocks.
- Study the Plotly curves after every simulated epoch.
- Repeat with another batch size while keeping other settings fixed.
- Rank results using speed, stability, convergence, accuracy, and memory.
Frequently asked questions
Is the training real?
The game uses a deterministic educational simulation. It models common trends without training a production model.
Why can results vary?
Noise, seed, optimiser, learning rate, and model complexity influence every simulated trajectory.
What is full-batch training?
One update uses the complete training set. It is smooth but can be slow and memory intensive.
What is convergence?
Convergence means loss improves slowly and approaches a stable level or chosen target.
Which batch size is best?
No universal winner exists. Hardware, dataset size, model architecture, and accuracy goals determine the best choice.
Why measure throughput?
Throughput shows processed samples per second. It separates computational speed from epoch count.