mhough/inceptionv1-tuning-atlas

InceptionV1 tuning atlas — every unit, with its nulls

Quantitative per-unit orientation and curvature tuning for all 5,808 units of InceptionV1 (torchvision googlenet), each shipped with two null distributions.

Why this exists. The Distill Circuits thread established oriented-edge and curve detectors with feature visualizations and rendered tuning-curve widgets. None of it was published as numbers, no randomization control appears anywhere in the thread, and OpenAI Microscope — the visualization layer — has returned HTTP 503 since roughly January 2025. The founding rung of mechanistic interpretability is currently its least reproducible.

Fields (one JSON object per unit)

layer, unit, osi (orientation selectivity, 1 − circular variance, measured at that unit's preferred spatial frequency), pref_orientation_deg, pref_curvature, curve_selectivity_index, response_magnitude, and per-null: null_osi_{mean,std,max}, osi_z_vs_null, osi_exceeds_all_nulls, null_alive, shuffle_osi_{mean,max}, osi_exceeds_all_shuffles, shuffle_alive.

Two nulls, because the standard one is too weak

• —Random-init (Adebayo et al. 2018): same architecture, random weights.
• —Weight-shuffle: permute each trained kernel's weights within-channel — preserves the weight distribution and keeps activations alive.

Both are distributions over 5 seeds. The distinction matters:

A randomly-initialized network is nearly dead below inception3a (0% of units respond), so "beats the random-init null" is trivially true there. null_alive flags exactly where the weaker null is uninformative.

Honest headline

Trained top OSI 1.000 vs random-init 0.528 and shuffle 0.980 — the sharpest detectors are real. But the median conv1 unit sits at ~0.25: "conv1 is all Gabor filters" is too strong a reading, and the numbers say so where the pictures could not.

Method traps this cost us

1. 1.Measure orientation tuning at each unit's preferred spatial frequency — averaging across frequencies inverted the conv1 result entirely.
2. 2.Use a bounded curve index, not a ratio; a near-zero straight-line response produced "selectivities" of 2,000,000.
3. 3.A dead null is not a null.

Caveat on unit indices

These are torchvision `googlenet` channel indices. They are not known to correspond to the unit numbering in the Distill articles, which refers to the lucid/TF-slim InceptionV1 checkpoint. Do not join these to those labels without first establishing the mapping.

Reproduce: scripts/inceptionv1_atlas.py (CPU, minutes). Interactive viewer and curriculum: spinning-up-in-mech-interp. Morgan Hough, Orthogonal Research and Education Lab (OREL).