Make learning visible.

Shape the data, build the layers, and watch a neural network learn in real time.

Epoch 000,000 Ready to train

Data

Pattern

Circle
XOR
Gaussian
Spiral
Draw your own
Plane
Multi-Gauss
250 train250 test
500 samplesseed 0.68936

Network

361 parameters
5 hidden layers
Click to edit this value
Weight0.2
7 inputs, 1 outputpositive weightnegative weightDashes flowing back show each weight’s next change: violet up, amber down.
Click an input to switch it on or off. Hover a neuron to see what it has learned. Click a connection or bias to edit it.Swipe to see the whole network

Output

Training loss—
Test loss—
Loss over time0 epochs
Run training to draw the loss curve.

Decision boundary

Output
−1+1

Understand every connection.

No black box. Just data, a few neurons, and a pattern waiting to be found.

Start with the data.

The dots are examples. In classification, their colors are the two categories your network is learning to separate. In regression, the goal is to predict a continuous value.

Training examples teach the network. Test examples show how it performs on data it has not trained on.

Follow the signal.

Each neuron combines its inputs using weights and a bias, then applies an activation function. Each layer transforms the signal a little further.

Training adjusts the weights to reduce prediction error. Hover over any neuron to see the pattern it has learned.

Read the colors.

Violet means positive. Amber means negative. Thicker connections have larger weights. The small square beside each neuron shows its bias.

The output background shows predictions across the space. Color intensity shows the output’s magnitude, not calibrated probability.

Set up a classroom view

Choose which controls your students see.

Switch off anything that would distract from today’s lesson. Apply the view here, or open it in a new tab to share the starting configuration.

Open this viewVisibility changes take effect on reload.

Keep following your curiosity.

Go deeper with Neural Networks and Deep Learning by Michael Nielsen, or Deep Learning by Ian Goodfellow, Yoshua Bengio, and Aaron Courville.

Open-source foundations and credits

Neural Network Visualizer is an independent studio interface built on the educational neural-network engine from TensorFlow Playground, created by Daniel Smilkov and Shan Carter. Original work copyright 2016 Google Inc., used under the Apache 2.0 license. It is not an official Google or TensorFlow product.

The original project builds on Andrej Karpathy’s convnet.js demo and Chris Olah’s neural-network explanations, with contributions and guidance from D. Sculley, Fernanda Viégas, Martin Wattenberg, the Big Picture team and Google Brain. This small engine is for exploration; TensorFlow supports larger applications.

Share this experiment

Copy this address to share the starting configuration and random seed.

The link does not include trained weights or the current epoch.