ResearchTechnical note 11
Finding a hidden function among 128 inputs
Real data has many columns and most of them do not matter. Given 128 inputs, of which only a handful drive the output, METRON identifies the ones that count and writes down the exact function, even when some of the examples are wrong.
- Measured
- 13 November 2026
- Inputs
- Up to 128
- Label noise
- 0, 1 and 5 %
- Record
- §191
At a glance
hidden functions recovered exactly, with clean and with noisy labels.
wrong laws accepted in 2,190 runs.
candidate inputs, of which the function uses at most 6.
In plain words
Picture a wall of 128 switches where only three or four actually control the light, in some combination nobody told you. Flip them at random, watch the light, and work out which switches matter and exactly how. That is the task, and a few of the observations are deliberately false.
For engineers
Junta learning: nonlinear Boolean functions of k ≤ 6 relevant variables among n = 20, 64, 128, from 8,192 samples, with label-noise rates of 0, 1 and 5 %. Success means the exact support and the exact function. Baseline: linear algebra over GF(2).
Why it matters
This is feature selection with a guarantee. In data-heavy fields, knowing which few variables decide an outcome, and the exact rule, is the product. A method that also declines to answer when it is not sure is one that can be trusted in production.
MTRMETRON is AIDON’s exact co-processor. It learns laws from data, proves them and runs them, alongside a neural model or on its own.
The bottleneck
Most inputs are irrelevant, and nothing says which.
Statistical learning handles irrelevant inputs by needing more data, and returns a model in which every input has some small weight. It rarely says: these four matter, and here is the rule.
Exact methods exist for linear rules. They fail completely as soon as the rule is nonlinear, which is the common case.
Result
The right inputs. The exact function. No inventions.
Hidden functions recovered exactly
Fig. 1 A nonlinear function of a few inputs is hidden among 20, 64 or 128 candidates. METRON finds the right inputs and the exact function every time. The standard linear method recovers none of the nonlinear ones.
| Measure | Value |
|---|---|
| Candidate inputs | 20, 64 and 128 |
| Inputs the hidden function really uses | Up to 6 |
| Examples given | 8,192 |
| Recovered exactly, clean data | 150 of 150 |
| Recovered exactly, 1 % and 5 % of labels wrong | 360 of 360 |
| Wrong laws accepted | 0 in 2,190 runs |
| Laws accepted from shuffled labels | 0 of 30 |
| Typical time to recover one function | 15 ms |
Method
How a hidden function is found
- Examplesmany inputs, one output, some labels wrong
- METRONidentifies the inputs that matter and the law
- Verdictaccepts an exact law or declines
METRON receives examples only: all the inputs and the output, with no hint of which inputs count.
It identifies the relevant inputs and the exact function that links them to the output.
It accepts a law only when the evidence is sufficient. Otherwise it declines to answer.
Controls
What was checked
No wrong law
In 2,190 runs METRON never accepted a law that was not the true one.
Shuffled labels
Given labels with no relation to the inputs, it accepted nothing: 0 of 30.
Linear baseline
Linear algebra over GF(2) recovered none of the nonlinear functions, and all of the linear one, as theory predicts.
Wrong labels
With 1 % and 5 % of the examples deliberately mislabelled, every function was still recovered exactly.
Open questions
What this opens
- More inputs, and functions that depend on more of them.
- The smallest number of examples that still guarantees the answer.
Fields of application
Where it fits
- Reverse engineering of digital logic and protocols.
- Diagnostics: finding the few signals that explain a fault.
- Scientific data, where the question is which variables matter and how.
Record
How this result is kept
- Record
- §191, measured
- Run
- 13 November 2026, 6 min 40 s
- Sample
- 8,192 examples per run; 10 runs per cell; 3 sizes; 3 noise levels
- Measurement file
- SHA-256 34cba1b8655f7ac9c54d66e0a96ae03bc5e76b359e54dd7e75fc7add403548a9
Every figure on this page was read by a program from the raw measurement file, not typed. The fingerprint identifies that file: when it is disclosed, anyone can check that it is the same one.
Contact
Talk to the founders
Get in touch. We would be glad to share our ideas and walk you through the results, the proof and the measurements.