Connections shape
behavior.
The cursor sees only pixels. Signals pass through a model of neural populations and return to the environment as motion and clicks.
Based on the cell-type diagram of a small octopus vertical-lobe volume in Bidel, Meirovitch et al., 2023. This is a population model with engineered equations and adapters, not a reconstructed brain.
A synthetic raster of the target's color and shape, without page text. Target coordinates are not given to the controller. Eight directional channels are computed from pixel colors around the cursor. Channel projection and the motor decoder are engineered adapters.
Disable the connections.
Compare the result.
Each condition uses the same targets, starting position, and time limit. Shuffling does not have to reduce performance: we show the measured result.
No test has run yet. Results are not prefilled.
| Condition | Successes | Time to success* |
|---|---|---|
| Run the comparison | ||
An inspectable source.
Directions and annotated synapse counts between seven cell types were manually transcribed from Figure 2. The authors show edges with more than 10 annotated synapses. Smaller connections and individual-neuron wiring are not reconstructed here.
Open the original Figure 2 ↗Download the graph and assumptions ↓Bidel, Meirovitch et al. · eLife 2023 ↗The limits of the model.
- Seven population nodes are repeated in eight engineered directions: 56 dynamic variables, not 56 biological neurons.
- Activity equations, time constants, weight normalization, vision, and the motor decoder are engineering assumptions.
- Some connection signs follow the authors' hypotheses rather than measurements of every connection. This models SAM/CAM balance, not complete physiology.
- This controller does not learn. The aquarium jar is a separate Q-learning experiment.
- Reaching a colored button does not establish page understanding, superiority of the biological diagram, or the ability to use arbitrary websites.