Neural compass

Neural compass#

This experiment builds a spiking ring attractor that stores heading as an activity bump, integrates a dark turn, and then maps recovery after a sector of the ring is silenced.

Prompt#

Build a ring of head-direction neurons that holds an activity bump like an internal compass. Point it north, rotate the animal in darkness, and show whether the bump follows the turn. Then silence a wedge of the ring and test every starting direction to discover when the compass recovers or fails.

Agent decision path#

  1. Classify the mechanism as a point-neuron recurrent attractor.

  2. Route neurons to brainpy-state and delayed binary communication to brainevent.

  3. Construct 48 head-direction neurons with dense ring connectivity.

  4. Encode angular velocity as a signed bias on recurrent transmission.

  5. Cue north, remove the landmark, and integrate a 90 degree/s dark turn.

  6. Silence a 60 degree wedge and preserve matched intact controls.

  7. Map all 48 starting headings with state-aware vmap2 inside the time loop.

  8. Classify spared, recovered, and failed states from error, coherence, activity, and departure.

Result#

The bump moves 94.9 degrees during a commanded 90 degree turn. Across 48 starting headings after the lesion, 11 are spared, 4 recover, and 33 fail.

Neural-compass panels show an activity bump shifting around a ring during a dark turn and a lesion outcome map over all starting headings. Most headings fail after the wedge is silenced, with four recovered cases.

Fig. 11 Continuous heading error and bump coherence support the recovery labels rather than leaving them as visual judgments.#

With-BrainX skill/Without skill comparison#

With BrainX skill

Source captured · benchmark pending

Preserve all headings, matched controls, delay checks, and recovery predicates when measuring execution and code generation.

Without BrainX skill

Matched run not collected

Compare only after the baseline reproduces the same lesion protocol and continuous metrics.