Layer 4 Circuit Diagram

This diagram is the primary output of the circuit extraction pipeline for a fine-tuned Faster R-CNN pot detector. It visualizes the minimal co-activation circuit in ResNet Layer 4 — the 2048-channel bottleneck where pot-specific features are most concentrated.

Each node represents a layer4 channel ranked by ablation importance score: how much detection mAP drops when that channel is zeroed out on a held-out tile set. Edges represent co-activation strength above a 0.3 Pearson threshold, measured across 64 orthomosaic crops. The layout is force-directed, so clusters of tightly co-activating channels naturally pull together.

Use the Plotly toolbar to zoom, pan, and hover over individual nodes for channel index and importance score. The diagram is generated from circuit.npy and channel_importance.npy produced by the extraction notebook.

Scroll horizontally within the viewer on smaller screens. Zoom and pan controls are in the Plotly toolbar (top-right of the diagram).

Reading the diagram

Node size scales with importance score — larger nodes are the channels whose ablation causes the biggest mAP drop. The top-5 channels by importance form the circuit's core; removing them together drops detection recall by ~40% on the test set.

Edge weight (line thickness) reflects mean co-activation across crops. Dense clusters typically correspond to spatially adjacent feature detectors that fire together on pot-shaped objects; sparser cross-cluster edges indicate feature combinations that matter for discrimination against similar objects (bowls, cups).