Gradient Descent Playground

Start here

1. Click the chart to add a few data points. Drag a point to move it; right-click one to remove it.
2. Move the w slider, or scroll while over the chart or slider, to change the model.
3. Try to make the red/orange error lines and the SSE score as small as possible.
4. Press Find Best w to compare your answer with the best fit for the current points.

More detail

• What the model does: the bias is fixed at zero. In Linear mode, the prediction is y = w·x. In tanh mode, it is y = tanh(w·x), which keeps predictions between -1 and 1.
• Residual lines: each red/orange line joins a data point to its prediction. Its length is that point’s error.
• Squared error: the bar chart shows each error squared; the SSE is their sum. Use the log-scale option to make very small errors easier to compare.
• tanh insets: in tanh mode, a small curve beside each point shows the pre-activation value w·x and where it lands after tanh.
• Sharing: the URL updates automatically with your data points, selected activation function, and slope. Copy it to share the same experiment.
• Scratch pad: notes are saved only in this browser, on this device.