Learning to ask without a history dictionary → · New result: learned spike thresholds →
Two different hidden light patterns illuminate the same wall. Without a useful occluder their signals are almost identical. Edit the mask or ask for a discriminating lens. This is a 1-D geometric explainer, not a real-camera result or a reproduction of the Python benchmark.
World A World B. Which one is behind the corner? The camera cannot see them directly.
Separation / sensor-noise norm: —. Larger means the wall can better distinguish the two worlds at fixed noise.
Gold cells block rays at an intermediate plane. The mask changes the linear measurement operator A, not the hidden object. Click cells to redesign it.
With fixed known geometry, another mask can expose image differences that the unmasked view nearly erases. The selection button chooses a mask using both candidate predictions, without using an unknown truth label; this is an illustrative oracle candidate set. This little page has no inverse solver, no real camera access, no sensor calibration, and no learned AI. The tested model, randomized candidate trials, frozen gates and results are in the repository.
Different stimulus histories leave different traces in twelve branches, yet give the same initial soma sum. Choose which branches reach the soma and see what a receiver can recover from three noisy readings.
Each branch retains a different leaky trace of these inputs. The blue-green and amber lines are past stimuli, not spike waveforms.
The un-gated sum cannot distinguish these eight histories. The branch traces still differ. A gate can select a distinction before it is mixed away.
You can see the hidden history for explanation. The selector receives candidate predictions and previous readings; it never receives the true history label.
This diagram shows the current traces and selected sum before noise. Received noisy readings from completed gates appear below.
The message is a noisy scalar and the identity of the gate. No branch states are sent. An action-potential waveform is not modeled.
The observer's memory of previous readings shapes the next question. Its known candidate dictionary is a real prior and storage cost.
512 fresh history/noise/truth seeds per condition · 8 supplied candidates · 1 shared initial reading + 3 new readings · norm-matched gates.
| Strategy / control | Identified | Log loss ↓ | Original-answer error ↓ | State change |
|---|
Original-answer error is normalized prediction error on eight unused gates applied to the pre-read memory. Erasing memory still scores against the original history, so predicting zero cannot count as successful recall.
Frozen protocol · Results and limits · Machine-readable receipt
The abstract model supports a narrow mechanism: retained branch history becomes recoverable through selected somatic sums, and retaining previous answers modestly improves which gate is chosen next. Reads can also disturb that memory.
This uses linear leaky traces, a programmed gate selector, a correct model and a known eight-history dictionary. It does not establish a learned neural gate, spike-waveform communication, unknown-history recovery or a new AI architecture. Twelve branch values are accompanied by 96 candidate-state values, eight belief values and shared model/catalogue storage in the observer. Selection evaluates all 24 candidate gates; computation is not matched to random selection.
The original PerceptionLab ecg.json accident supplies the historical question about observation inside feedback. Its aliasing/variance-controller pulse has different equations. AnttisBrain2's moons, MovingTarget2 and BrainAsInverseModeler remain distinct contributors to this test.