Axon documentation
A guided manual for training, debugging, comparing, and exporting small neural networks in Axon Studio.
Onboarding
Axon Studio is a browser-based neural-network lab for small 2D classification problems. It is built to show how a model learns, not just what it predicts at the end.
First run
Open the playground, select a preset, press Run, then switch between boundary, errors, and gradients. Pause training and scrub the epoch timeline to inspect earlier states.
Datasets
Axon ships with four built-in datasets and supports custom CSV import. Pick a dataset based on the learning behavior you want to observe.
Step by step
Select the Circles clean preset, leave all defaults, and press Run. The decision boundary will start forming within the first few epochs. Use the epoch timeline to scrub back and see how the boundary evolved.
Power user
All major playground actions are accessible from the keyboard. These shortcuts work as long as focus is not inside a text input.
Visual model output
The central square maps the x₁ / x₂ input space. Purple and pink regions show the predicted class, while the contour line shows the 0.5 decision threshold.
Debugging mistakes
Errors view highlights points where the predicted label does not match the true label. This helps you understand whether the model is underfitting, overfitting, or confused by noisy data.
Point-level transparency
Click any data point on the boundary canvas to inspect how that specific input travels through the network layer by layer. The right panel shows activations, output score, and confidence.
Generalization
Axon reports train accuracy, test accuracy, and generalization gap. This stops the playground from becoming a pure memorization demo.
Layer-level learning
The Gradients view and layer flow indicators show how strongly each layer is updating. Use this to detect vanishing gradients or stuck layers.
Custom datasets
CSV mode supports two numeric input columns and one binary label column. Axon normalizes x/y values to [-1, 1] so the dataset fits the boundary canvas.
Auto-parsing
When you upload a CSV, Axon scans the header row and attempts to auto-detect which columns are numeric inputs and which is the binary label. You can override the selection manually.
Error handling
The CSV importer validates each row before training. Rows with non-numeric values, missing fields, or out-of-range labels are flagged and excluded.
Data preprocessing
All input features are normalized to [-1, 1] based on the min/max of the uploaded dataset. This ensures the values fit the boundary canvas and the model trains stably.
Data quality
Rows that fail validation are silently excluded from training. The importer shows a count of invalid rows so you can assess data quality before proceeding.
Experiment tracking
Save multiple runs from the playground and compare them on the Experiments page. Each saved run stores the architecture, dataset, metrics, and training config.
Portable results
Export a full experiment snapshot as JSON. The file includes metrics, architecture config, dataset metadata, snapshots, and saved runs.
Visual export
Download a screenshot of the current decision boundary canvas as a PNG. The export captures the exact visual state including points, boundary regions, and confidence zones.
Quick share
Copy a human-readable text summary of the current experiment to your clipboard. The summary includes dataset, architecture, epoch count, loss, and accuracy.
Run history
The Experiments page shows all runs you have saved from the playground. Runs are stored in localStorage and persist across page refreshes.