TensorFlow Playground: watch a neural network learn, free in your browser
Press play and watch a neural network learn to separate data, epoch by epoch, with every knob free to turn. No account, no install, no cost. Here is what the official pages confirm and what our hands-on test found.
Quick facts
- What: TensorFlow Playground, an interactive browser demo that lets you train a small neural network and watch it learn in real time
- Cost: completely free, no signup, no paid tier; open source under the Apache 2.0 licence
- What you control: dataset, learning rate, activation function, regularization, number of hidden layers and neurons, noise and batch size
- Made by: Daniel Smilkov and Shan Carter, with feedback from Google Brain teams; hosted in the TensorFlow GitHub organisation
- Runs on: any modern browser; the training happens on your own computer, nothing to install
- Privacy: no account; the page sets Google Analytics cookies (details below)
- Works in India: yes, global site, English only
What it is
TensorFlow Playground is a small website with one big idea: a neural network you can tinker with directly. You pick a dataset, press play, and watch the network reshape itself epoch by epoch until the orange and blue regions separate. Every important knob of a real neural network is there to turn: learning rate, activation function, regularization, the number of hidden layers and neurons, noise and batch size. As its header says: "Don't worry, you can't break it."
According to the credits on the page, it was created by Daniel Smilkov and Shan Carter, building on Andrej Karpathy's convnet.js demo and Chris Olah's articles, with feedback from the Google Brain teams. The source code is public in the TensorFlow GitHub organisation under the Apache 2.0 licence, with a public repository dating to April 2016 and over 13,000 stars. The site also links two free companion books for going deeper, "Neural Networks and Deep Learning" and the "Deep Learning" book.
What we tested
On 22 October 2026 we opened the site in a desktop browser with no account. This is exactly what we did and saw:
- The page loaded straight into the working demo: no signup, no download, no paywall.
- We pressed play on the default circle dataset. The epoch counter started climbing immediately, and the loss curves on the right began falling.
- After about 160 epochs, the output panel showed test loss 0.006 and training loss 0.005, and the background of the data panel had split into a clean orange ring and blue centre, meaning the network had learned to separate the two classes.
- Every control we touched responded live: pausing, changing the learning rate and adding hidden layers all took effect without any account.
Is it free?
Yes, entirely. There is no account system and nothing to buy. The code is open source under the Apache 2.0 licence, which even permits reuse of the code itself. It is an educational demo from the TensorFlow team, not a product with a premium tier.
What you can use it for
- Understanding how AI learns: seeing loss fall as epochs pass makes "training" concrete in a way no article can
- Feeling why settings matter: set the learning rate too high and watch the network fail; too low and it crawls
- Classroom demos: four datasets, a regression mode, and shareable URLs that encode your exact setup for students to open
- A stepping stone: the page itself points you to two free books when you want the maths behind the demo
Privacy and data
There is no account, so you give the site no name or email. The training runs entirely in your browser on your own computer. In our visit we observed the page setting Google Analytics cookies (_ga and _gid), which are used for visit statistics; the site does not publish its own privacy policy page, and the footer links only to its GitHub repository and licence. We did not audit the hosting servers; this section describes only what we could observe directly.
India
The site is global with no region restrictions and loads fine from India. It is English only. Because the neural network is tiny and trains on your own device, it works on modest laptops, though a desktop or large screen makes the graphs far easier to read than a phone.
Good to know
- It is an educational demo, not the full TensorFlow library: the site says it uses a tiny custom neural network library built for this visualization.
- The spiral dataset is the classic challenge: most simple setups fail on it, which is itself the lesson about why deeper networks exist.
- Your configured network is encoded in the page URL, so bookmarking or sharing the link preserves your exact setup.
- The project is stable rather than frequently updated; it has looked the same for years, and that is fine for what it teaches.
Frequently asked questions
Do I need to install anything or create an account?
No. It runs entirely in the browser and there is no signup.
Is this the same as TensorFlow?
It comes from the TensorFlow team and lives in their GitHub organisation, but it is a teaching demo with its own tiny library, not the full TensorFlow framework.
Is it suitable for beginners?
Yes. You can ignore every control, just press play, and still watch a network learn. The knobs are there when you get curious.
Can I use it offline or in teaching materials?
The code is Apache 2.0 open source on GitHub, so you can run a local copy; check the licence file in the repository for the exact terms.