Build a tiny brain. Watch it learn. Change the architecture, explore real data, and see how a decision boundary emerges with every epoch.
Real backpropagation5 datasetsLive decision boundaries
Input → Hidden layers → Output
INTERACTIVE WORKSPACE LOCAL
Train a model, explore a boundary Ready
Unrecognized saved workspace
The stored data was not replaced. Download a backup or explicitly reset your workspace.
Ready to trainEpoch 0
Training is idle
LIVE CLASSIFICATION
Decision boundary
Class 0 Class 1
Click the chart to probe any coordinate and see its predicted probability.
Changes predictions and holdout diagnostics only. Training history keeps the standard 0.50 threshold.
True positives—
False positives—
False negatives—
True negatives—
Holdout evaluation appears here.
Training accuracy—
Test accuracy—
Train loss—
Test loss—
● Runs locally in your browser. Your datasets and model are not uploaded.Educational binary classification · x and y inputs · 0/1 output
LEARN BY EXPERIMENTING
Change one variable. See what learns.
Choose a nonlinear dataset, add hidden neurons, and watch train versus test loss. Try fewer neurons, stronger regularization, or noisier data to explore underfitting and overfitting.
01
Choose a pattern
XOR, circles, clusters, spirals, and moons illustrate different decision boundaries.
02
Design your network
Set 0–4 hidden layers, up to 12 neurons per layer, and pick the activation and optimizer.
03
Train and inspect
Use genuine gradient updates, a test holdout, and interactive probabilities—not simulated results.
Built for education, not production machine-learning predictions. Training is nondeterministic across some browser/platform floating-point environments even with a fixed seed. A high training accuracy does not imply generalization. Model weights and history are local; saved JSON files may contain imported data. Canvas rendering can approximate the decision boundary between sampled grid points.
Using Neural Network Playground
Choose a supported dataset and network configuration, then train or step through learning. Compare the decision boundary and loss history, and change the prediction threshold to inspect holdout metrics. Export a model/project backup or a bundle of supported artifacts.
This is a small educational 2D binary classifier, not a production AI service or general model-training platform. It supports bounded datasets, layers, neurons and epochs. Threshold changes do not retrain weights or change fixed-threshold history curves. Metrics depend on generated data/settings; exported bundles can include full datasets and model history. Browser storage is best effort.
Workflow and output choices for Neural Network Playground
Experiment with a bounded 2D binary classifier using generated datasets and a small configurable network. Train or step through learning, then compare its decision boundary and loss history.
A project/model backup or artifact bundle preserves supported settings and results. Bundles may include complete generated datasets and training history; review their contents before sharing.
Example workflow
Choose one of the supported generated datasets and a small network configuration.
Train or step through learning with a supported optimizer, then inspect the boundary and loss history.
Change the prediction threshold to compare holdout metrics and export a project or artifact bundle for later review.
Authored example: Compare a simple generated two-class dataset using Adam and SGD with the same small network settings. Record what changes in the boundary and history without treating one run as evidence of general model performance.
Limits and handling: This is an educational 2D binary-classification lab, not a production AI platform or general model-training service. Dataset, layer, neuron and epoch bounds apply; threshold changes do not retrain weights or rewrite fixed-threshold history curves.
Neural Network Playground questions
Can I train or deploy a production model with this playground?
No. It is a bounded educational 2D classifier, not a general training or deployment platform.
Does changing the prediction threshold retrain the network?
No. It changes prediction and related holdout metrics; it does not retrain weights or rewrite fixed-threshold history curves.
