MICrONS

BRAINS TEACH MACHINES / NAVIGATION STUDY 01

Branching monochrome neurons surrounding the MICrONS wordmark
CORTICAL NAVIGATION / STUDY 01
SHARED TRACE / OBSERVATION ONLY
CORTICAL NAVIGATION INSTRUMENT

A brain learns
the shape of escape.

An unfamiliar maze is introduced every 210 seconds. The navigation model must probe, hesitate, reverse and recover without access to the solution path.

01 / CORTICAL ACTIVITY

Functional field

DORSAL ORIENTATIONINTERACTIVE VOLUMEDRAG · ORBIT · INSPECT
CELLS200,000dataset scale
PHYSIOLOGY75,000neurons
SYNAPSES523 Mdetected
FIELD RATE38.6Hz mean
02 / SPATIAL INFERENCE

Trial 0001

SEARCHING
NEXT MAZE03:30
STARTACTIVE TRACEREJECTEDGOAL
ELAPSED00.00seconds
WRONG TURNS0branch probes
EFFICIENCY—vs optimum
EXPLORED000cells
SERVER-CONTROLLED0.65–1.15 S / STEPJUNCTION DELIBERATION ENABLED
SUBJECT / CURRENT RUN

Mouse P87

TOPOLOGY—maze fingerprint
REVERSALS0steps retraced
DECISION STATESEARCHcurrent operation
MEMORY LOAD0%active trace
TRIAL WINDOW210 Sglobal cadence
PERFORMANCE ARCHIVE
FASTEST SOLVE—SECONDS
AWAITING
BASELINE
COMPLETED0
FAILED0
MEAN TIME—
SUCCESS RATE—
TRIAL TRACE / LEARNING LOG

Encoded observations

OPEN RAW FEED ↗
T—INIT

Spatial prior initialized. Exploration weight set high.

α 0.720
CURRENT INFERENCE STATE

Inside the decision

PLASTICITY INDEX0.720adaptive gain
ROUTE ENTROPY0.00bits / junction
WORKING MEMORY0%path occupancy
FAILURE HAZARD18%per trial
DOCUMENTATION / PROTOCOL 01

From connectome
to navigation trace.

This record describes what the instrument measures, how a maze trial unfolds, and how to read the live cortical and behavioral signals without confusing dataset-scale anatomy with the navigation model’s runtime state.

01 / RESEARCH QUESTION

Can a circuit-informed agent build a useful spatial prior from repeated uncertainty?

The study observes navigation under imperfect information. At the beginning of every trial, the model receives only the current open corridors. It does not receive the goal path. Progress therefore emerges through local decisions: advance, compare branches, commit, detect contradiction, reverse and resume.

Success is not defined by a straight path. A credible trace contains uncertainty, discarded hypotheses and recoverable error.
02 / STRUCTURAL INPUT

The anatomical reference is a cubic millimetre of mouse visual cortex.

The MICrONS public release spans all six cortical layers and includes roughly 200,000 cells, functional recordings for approximately 75,000 neurons and more than 523 million detected synapses. The volume combines two-photon functional imaging, microCT alignment and serial-section electron microscopy [01][02].

MICrONS diagram connecting functional recordings, electron microscopy reconstruction, neurons, synapses, and functional connectomics
FIG. 01

Functional measurements and structural reconstruction meet at the level of cells, processes and synapses.

MICrONS TUTORIALS [01] ↗
VOLUME1.4 × 0.87 × 0.84 mm
CORTICAL LAYERSL1—L6
FUNCTIONAL CELLS≈75,000
DETECTED SYNAPSES>523 million
03 / TRIAL CYCLE

One shared clock controls every observer.

  1. Encode topologyA seeded 21 × 21 corridor graph is generated on the server and assigned a reproducible topology hash.
  2. Search locallyThe route advances one cell at a time. At multi-branch junctions, the trace pauses while competing continuations are evaluated.
  3. Reject and reverseDead ends and low-confidence branches remain visible as rejected evidence. The model retraces to the last viable junction.
  4. Consolidate outcomeA goal or failed trace is written to the record with elapsed time, explored cells, reversals, efficiency and outcome.
  5. Reset globallyAt the next trial boundary, all viewers receive the same maze, trace and state from the shared server clock.
04 / SIGNAL GLOSSARY

Four measures expose the decision process.

PLASTICITY INDEX

Adaptive gain applied when recent evidence should alter the next route comparison.

ROUTE ENTROPY

Uncertainty across available continuations. It rises at ambiguous junctions and falls after commitment.

WORKING MEMORY

Occupancy of the active place-field trace, including the current hypothesis and recoverable route history.

FAILURE HAZARD

Protocol-level chance that a trial terminates before the goal after memory or confidence destabilizes.

05 / LEARNING RECORD

Every wrong turn remains part of the evidence.

The event stream separates route advance, junction deliberation, branch commitment, exploratory probes, reversals, recovery and final encoding. The archive keeps completed and failed trials together so performance records cannot erase unsuccessful search behavior.

THINK compare continuationsPROBE enter an uncertain branchREVERSE unwind a rejected pathENCODE consolidate the outcome
06 / DATA PROVENANCE

Source material and reproducibility.

Dataset access patterns, annotation concepts and programmatic tooling are documented in the MICrONS tutorials. The cortical-mm³ release describes the imaged volume and reconstruction pipeline. Functional-connectivity analysis methods and figure reproduction are available in the companion 2025 repository [03].

[01]MICrONS TutorialsDataset access, annotations, connectivity and Python tooling↗ [02]Cortical cubic millimetreImaging volume, reconstruction, physiology and inventory↗ [03]Functional connectivity repositoryMethods and reproducible analysis for mouse visual cortex↗