—
Threshold
−0.150+0.15
04 / DICTIONARY-FREE BINARY QUERIES
Learning to ask
with a spike.
The past remains in twelve branches. A receiver learns which branches to listen to and where to place a threshold. Each answer is one bit; previous answers can shape the next question.
Actual trained weights run in this browserFrozen protocol · 3 initializations
Additional reads0 / 8
Parameters, every control1,252
Independent test histories512 / condition
Main MSE gain vs fixed—
01 / ASK THE MEMORY
Same past. Different questions.
Both receivers use the same example history and sensor-noise stream. The history is shown for inspection; the receivers get only replies and their own commands.
—
Threshold
−0.150+0.15
The first faded answer is a common, nondestructive initial read. Only the following eight count toward the budget. Gates have unit Euclidean norm. The threshold changes the bit; the write changes branch memory.
02 / ASK SOMETHING NEW
What did the replies preserve?
These signed questions were withheld as a question family during training. They become available after read eight. Changing a final question uses the same receiver state and asks no new physical question.
Selected receiver—
Comparison receiver—
Original memory's answer—
Absolute errors—
Original branchesSelected receiverComparison receiver
Complete the reads to reveal the final forecasts.
03 / HELD-OUT EVIDENCE
The result has to earn its claim.
512 histories, eight sensor-noise trajectories per history and all three trained initializations. Intervals resample whole histories, conditional on these trained models. No test-based tuning.
| Receiver | Answer NRMSE ↓ | Full loss J ↓ | Squared displacement ↓ |
|---|
NRMSE 1 equals predicting zero. Graded references transmit a scalar rather than one bit: same read count, different information per answer. Their accuracy comparison does not establish an equal-bit advantage.
Frozen protocol · Outcome ledger · Numerical receipt · All 24 weights and training traces · Inference verification
WHAT WE CAN SAY
A learned interface to retained history.
What is actually matched
Every control has the same recurrent network and internal gate-and-threshold head, even when that head does not control the plant. Each gets 1,252 parameters and 7,376 counted neural MACs per episode. Sender and receiver retain 28 values.
Adaptive commands send thirteen coefficients per read; fixed schedules can be shared in advance. Counting excludes nonlinearities and plant work. Equal counts do not guarantee equal effective capacity.
Where this question came from
The original ecg.json feedback accident, the moons in AnttisBrain2 and Varjoluotain's mask put observation geometry inside the problem. Kuulustelu made the threshold's role experimentally explicit. Here the commands are learned.
This remains a simulation with known training dynamics, correlated non-Gaussian histories and an externally controllable threshold. The invertible write moves memory. It does not prove biological spiking, a new neuron or practical AI superiority.