BUILD A NETWORK.
THEN YOUR OWN AI.
First, write the rules. Then let a network learn its own.
Eight experiments with a tiny synthetic flight logbook.
You write the rulebook.
Choose a condition and how many minutes it adds. Start with two rules from the app. Change a threshold, then inspect a flight just below it. Can your rulebook beat guessing everyone is on time?
Your delay recipe
Up to 3 rulesMatching rules add together. Conditions use the original clipboard units. The native beginner starts at zero: wind above 30 mph adds 60 minutes; rain above 2 inches adds 30.
Just under the threshold.
8 facts make 28 pairs and 56 triples. A few thresholds can miss in-between cases and combinations. This example stays fixed while you edit the rules.
Pack your clipboard.
A network can only learn from the facts you give it. Keep useful clues, try the shoe-colour decoy, and see how raw numbers become comparable.
Try this: leave the shoe-colour decoy on, inspect the wind dial, and compare its raw value with its distance from an ordinary practice flight.
Inputs for your network
Changing the clipboard rebuilds the network and resets its training scores. These choices carry into lessons 3 and 4.
Give each fact a route.
Each input contributes separately. Train the weights and the bias, then compare the practice and quiz errors. What happens when a holiday weekend and a nearly full plane need to combine?
Try this: train 100 steps, then inspect the compound flight. The table at the gate shows exactly how each direct route adds its share.
Each route adds its share.
Actual contributions / minutes| Route | Raw fact | Standardized | Weight | Minutes |
|---|---|---|---|---|
| Their sum / the network guess | ||||
Contributions include conversion from standardized delay back to minutes. The starting offset includes the practice-average delay and learned output bias. Displayed rows are rounded.
Direct routes are additive: rain has the same coefficient regardless of the wind. They cannot create a rain × wind interaction. Hubs let facts combine before reaching the gate.
Build a layover.
A hub combines its incoming signals through a ReLU gate. Add a row, change its size, and let the network learn all its weights. More hubs give it more possibilities, not a guaranteed better quiz score.
Changing the layout restarts training.
Try this: train one row, pause, then add a second row and train again. Your paused scores are kept below. More rows need evidence, not faith.
Keep the evidence.
Quiz MAE / minutes / smaller is closer| Approach | Quiz error | When measured |
|---|
A model score is captured when training pauses or before changing its layout. Clipboard changes clear model comparisons, so these scores always use the same inputs and quiz. Rules use raw facts independently.
Which hub follows the storm?
Train a hub model, then check how first-row activations associate with the synthetic storm factor across all quiz flights.
A correlation is a clue, not proof that a hub “means storm.” More hidden rows make a deeper network; better quiz performance must be measured.
Feel your way downhill.
This is a measured slice of a network’s practice loss. Only two selected parameters move. Follow their gradient with small steps, then try a larger step.
What wakes this hub up?
A hub card compares activity with the synthetic flight factors, shows contrasting flights, and tests what changes when the hub is closed. A correlation alone does not establish a meaning.
Try this: inspect a hub, give it a provisional name, then reveal what its card supports. Quiet and mixed hubs are valid findings.
What wakes it up?
Pearson r: right means busier as a factor increases; left means quieter. Zero means no measured linear association; constant signals also receive zero in this card. All measurements use the quiz pile.
“Mostly” needs a strong, distinct association and a measurable loss when closed. Your name stays yours; it does not become a scientific label.
Open it only once.
The quiz helped you make choices. This fresh exam stays unscored until you open it. Compare your hub-model snapshot with an 8 + 8 hub crammer that practices a small pile repeatedly.
Try this: prepare the two candidates, let the 8 + 8 hub network cram, then open the exam. Read the measured results even if they surprise you.
Prepare a comparison to freeze your current hub model. The crammer will use the same inputs, two rows of 8 hubs, and only 64 practice flights.
| Candidate | Homework MAE | Exam MAE |
|---|---|---|
| Exam scores remain sealed. | ||
A low homework error can coexist with a large unseen error. More capacity does not guarantee either outcome.
The web uses a smaller synthetic logbook and its own training engine. Study-pile size differs; capacity may differ depending on your model. This comparison does not isolate the effect of network size. These are fresh exam draws, not the native app’s historical year splits. Results freeze on the first opening: tuning a model after seeing its exam score cannot earn another untouched test here.
Take a route out of service.
Close a connection or a hub, then compare exactly the same quiz flights. This is ablation: a temporary intervention on a frozen network. Nothing retrains.
Try this: close the shoe routes, reopen them, then close the hub most associated with storms. Does either change help, hurt, or do little on this model?
Click a route or hub to toggle it. Green solid routes = positive weights; blue dashed = negative. Gray dotted parts are closed. Green / blue node fill shows positive / negative activation; brightness shows magnitude. Use the selectors below for the same controls.
Got worse the most
Got better the most
Ablation measures a part’s effect in this network on these examples. It does not prove a unique concept or explain every prediction.
A little context
for your experiment.
This is a small, real learning system built for this page. It trains locally in your browser on a seeded synthetic logbook. It follows the NeuroScope AI airline lessons and their synthetic flight world, with a smaller logbook and a separate browser training engine.
Practice flights teach the network. Quiz flights give feedback on unseen examples. Repeatedly choosing a model by its quiz score can overfit that quiz, too. These scores say nothing about real aviation prediction.
The macOS course and its larger-model journey are separate. Explore the course ↗