01
Finds what a question touches
Activation spreads from the words of a question through the graph, two ties out, and the answer shows the part of the brain it used and what else the question bears on.
The brain
A knowledge graph that learns
Everything the instrument holds, joined up: where every figure comes from, the countries, blocs and treaty groups, the levers, the channels they work through, what they move, the events a sentence can name, the live feeds and your own files. Two neural networks are trained on it in your browser, and their hidden units and weights are part of the same graph.
How it is built
Four things, one structure
The graph and the networks are not two features side by side. A network is a graph whose ties have weights, and here it is drawn as one.
-
01 / 04
The graph as it is known
Nodes for every source of a figure, every figure, every country, bloc and treaty group, every lever and the channel it works through, every outcome, every event a sentence can name, every live feed and your company's files. Ties for where a figure comes from, who is aligned with whom, who belongs to what, who trades with whom, which channel a lever works through, and what it was measured to move.
- Provenance, alignment, membership and trade ties
- Levers measured on the engine one at a time
-
02 / 04
Two networks learn it
Network A learns how the world's figures hang together: from a country's size, income, age, structure, energy, capability and alignment to its stability, inflation, unemployment, life expectancy and governance. Network B learns how the simulation behaves: from the same figures to what the baseline run does with each country. Both are trained in your browser, seeded, so the same data trains the same weights anywhere.
- Ten and twelve hidden units, full batch Adam with weight decay
- Scored by threefold cross validation, on countries each network did not see
-
03 / 04
A country runs through it
Choose a country and its own figures travel through the network as light, in proportion to what each tie carries. The hidden units light up by how strongly the country drives them, and the countries whose hidden activations sit closest to it are the countries the network treats as alike across every figure at once.
- Peers by likeness of hidden activation
- Every node opens to what it holds and where that came from
-
04 / 04
Why it sits where it does
The network's prediction for one country is split into what each of its figures adds against a country of average figures, by integrated gradients. The parts add up to the difference exactly, each part carries the provenance of the figure behind it, and the fit on unseen countries is printed beside it so the reader knows how far to trust it.
- Exact to a rounding error, checked on every draw
- A weak fit is said in words, not buried in a score
What the brain does for the analyst
02
Explains a country
Integrated gradients through either network split a prediction into what each figure adds, with each figure's source.
03
Says what is alike
Countries alike across every figure at once, by the cosine of their hidden activations, rather than alike on one figure at a time.
04
Knows what a lever does
Every lever is run alone on the engine at its own default size, one path, six years, and what it did to growth, prices, jobs, output, stress and energy is kept as a tie.
05
Knows what it is unsure of
Every output of every network carries its fit on unseen countries, and the analyst will not lean on one that fits poorly.
06
Ranks what to upload
The figures the networks lean on hardest that are least firm are the ones whose measurement would sharpen the most answers.
The limits, said plainly
A pattern is not a cause
Two hundred and thirty eight countries are enough to learn what goes with what, and not enough to learn why. The networks are read that way everywhere they appear.
Small on purpose
A network this size can be opened and read, node by node. A deep one trained on two hundred and thirty eight rows would learn the countries rather than the pattern and could not be read at all.
Only as good as the record
A network trained on patterned estimates learns the pattern that made them. Every explanation says which of its inputs are estimates, and the data room says which to replace first.
Open the brain and click a node
It trains when you open it, in your browser, in a few seconds.
Open the simulator