HELP

Train starts the loop and stops it again, Step computes exactly one epoch, and Reset rolls fresh weights and empties the loss curve. The data stay as they are — new points come only from New sample.

Every tile shows how that neuron colours the plane: green for a positive output, orange for a negative one, pale where it is undecided. It is not its state for one particular point but its output for every place in the plane.

In the first layer these are mostly plain light-to-dark gradients; further back the patterns grow more tangled, because that is where the partial answers of the earlier layers are combined.

The colour is the sign of the weight — green takes the signal as it comes, orange turns it around — and the thickness is its magnitude. Thick lines carry the decision, pale ones carry almost nothing. In the browser the mouse pointer on a line shows the weight to three decimal places, and on a neuron its bias.

A click or tap on the line severs it; it is drawn dashed, transmits nothing and is no longer trained. A second click puts it back.

It is the most instructive experiment in the app: cut a thick connection while the training is running and the loss jumps up, and you watch in real time as the remaining neurons redistribute the work among themselves.

Filled dots are training points — they were learned from. Rings are test points, which the network never gets to see; they appear only with the switch Show test points, and a legend under the surface then names both shapes.

Not when every point is right, but when the boundary is smooth and the rings sit as well as the filled dots. A network that has grasped the rule behaves sensibly even where there are no points at all. Tongues and islands around single points are the opposite.

The learning rate is the step size: too small and nothing moves, too large and the loss jumps. The activation supplies the kink, without which even ten layers add up to a straight line. Regularisation penalises large weights. The batch size says after how many points the weights are adjusted.

Noise and training points concern the data, not the network — both draw a fresh sample.

The loss barely moves: raise the learning rate or add neurons. The loss jumps: lower the learning rate. The surface stays a straight line: the activation is linear, or a hidden layer is missing. All tiles pale: dead ReLU neurons, so reset with a smaller learning rate. Test worse than training: overfitting, so fewer neurons, more points or regularisation.

Section 13 of the manual has the same list as a table.

Because the noise displaces the points before the label is assigned. At high noise some points lie on the wrong side of the actual rule. That is the limit of the data, not of the network — and forcing it anyway means learning the noise by heart.

An epoch is one complete pass through all the training points. A batch is the portion of points after which the weights are adjusted once: small batches make the path jittery and quick, large ones calm and sluggish.

Because a different network then has to be built: layers, neurons, features, activation and data set all belong to that group. Learning rate, batch size and regularisation, by contrast, take hold mid-run without touching the epoch counter.

Largely yes. The network diagram, the decision surface and the loss curve say what is currently there — features, layers, accuracy, epoch and loss. The seven feature tiles are reachable one by one and toggle on a double tap. A label and its figure belong together and are read together: “Epoch 180”, not two separate stops. This holds for TalkBack on Android as for VoiceOver on iPhone, iPad and Mac.

What does not work: cutting a single connection. That would need two dozen targets nobody could tell apart, so it stays with a finger on the line.

On the very first launch an introduction of five pages lies over the app. If you skipped it or want to see it again: Help, and at the very bottom Show the introduction again. On Android the help sits in the drawer behind the menu symbol, on iPhone and iPad in the tab bar, on the Mac in the sidebar. By itself the introduction appears only once.

On the Mac and on an iPad with a keyboard, in the Lab menu: ⌘ ↩ trains and pauses, ⌘ . runs one step, ⌘ R resets, ⇧ ⌘ R draws a new sample, ⌘ T shows and hides the test points.

In the browser with the pair of switches DE/EN at the top right, beside the buttons. On a first visit the page follows the browser's language; after that the last choice applies, including after the window is closed.

In the app under Settings › Language: From system, Deutsch or English. It is independent of the language the phone is set to and survives a restart of the app. Either way it takes effect immediately, and the running training carries on.

Directly below it: From system, Light, Dark. Switching recolours the tiles, because the zero point of the colour scale sits on the tile colour — and in the dark that is a different one.

No. The network and the compute loop hang on the shell of the app, not on the lab screen; come back and the epochs have gone on counting. Rotating interrupts nothing either. Only when the app goes into the background on a phone or tablet does it stop. On the Mac it keeps computing, even with its window behind others.

The browser has the tooltip on hover, the apps do not, not even on the Mac — the line thickness remains the clue. Light and dark follow the operating system in the browser, while the app has a choice of its own for it. And the decision surface is coloured a little paler in the app, otherwise an orange point would vanish in an orange field. German and English are in both.

Everything is computed on the device. The Android app requests no permissions at all, internet included, so it cannot send anything; the app for iPhone, iPad and Mac asks for no permission either and opens no connection of its own. Both store only a handful of settings on the device: language, appearance, whether the introduction has been seen, and a launch counter for the one-off rating request. The browser version puts a single value into the browser's local storage, the language you picked; its only outward request fetches the fonts. The data points are generated afresh by a random generator on every start.

With the introduction: it explains neuron, weight, activation, loss, gradient descent and backpropagation with no prior knowledge, using diagrams and an exercise at the device. The manual takes those terms for granted.