Each model loads already trained, with notes on what to look for. The training history is kept, so you can scrub back and watch it learn.
Columns are layers, squares are neurons and curves are weighted connections. Signals flow left to right; error signals flow right to left during training.
Blue is positive and orange is negative everywhere: weights, activations, signals and the two classes of a binary problem. Violet measures gradients, the pressure to change.
One input is always being traced through the network. Hover over the Output map to move it, click to pin it, or click a data point to use that example.Drag on the Output map to move it, or tap a data point to use that example. The Neuron tab shows its arithmetic.
With Maps on, each neuron shows what it outputs for every possible input, so you can watch simple shapes in the first layer combine into complex ones deeper in.
Use − and + under a layer to change its size, the + between layers or the Hidden layers control to add depth, and the chip under each layer to pick its activation function. To give one neuron its own function, select it and use the Neuron tab. The chip under Output picks the loss function.
The Build your own button opens the model builder. Drag parts between the Input and the Output (or click them), drag blocks to reorder them, and drag one off the model to remove it; drop a Neuron on a dense layer to add it there. Click a block to change it, the Input block to choose or make the problem, and the Output block to choose the loss. Every time you train after a change, the run is listed under Runs with what changed. Architecture shows the model you built as a diagram you can zoom into, like every other model.
Training uses one part of the data. A second part, the validation set, is scored as it trains to show how well it generalizes. A third, the held-out test set, is only scored when you press Run held-out test, once you have finished tuning.
Click a neuron, or a part of any zoom view, to see what it is; double-click it to open it.Tap a neuron, or a part of any zoom view, to open it; hold a finger on it to see what it is. Zoom into any part of the network with the mouse wheel, a pinch or a double-tap, and drag to pan. On the output map and the hidden space, and in a window narrow enough for the page to scroll, hold Ctrl while you scroll. Labels stay readable and show full formulas once there is room. In a lab’s Architecture view, the mouse wheel or a pinch opens the part under it, and zooming back out returns a level.
Zoom far into a neuron and it opens: its weighted sum, its PyTorch code, what each library call does inside, single float multiplications and additions bit by bit, the adders and logic gates doing them, a gate’s transistors, and finally the electrons moving through a transistor’s channel. In the Architecture views every heatmap cell and bar opens the same way. Stored weights lead down through their bits to the memory cell and the electrons holding the charge.
On an iPhone or iPad, open this site in Safari, tap Share, then Add to Home Screen. On Android, open the browser menu and tap Install app. On a computer, use the install button in the address bar. The app opens full-screen and keeps working offline.