From ebc2e7f5129b81db745ad15e5b78b69a95b9837c Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Sat, 1 Aug 2026 14:11:25 -0500 Subject: Augmentation orbits do not substitute for multiple photographs Four random resized crops per Visual Genome image, segmented independently with their fields averaged, leave the correlation at 0.6559. Closed-world re-renders resample layout, which is nuisance by construction, so averaging removes modality-private variation; random crops perturb framing that self-supervised patch features already absorb, leaving nothing to cancel. Co-Authored-By: Claude --- LAB_NOTES.md | 14 ++++++++++---- NATURAL_RESULTS.md | 16 ++++++++++++++++ artifacts/vg_5k/natural_pipeline_orbit.json | 8 ++++++++ 3 files changed, 34 insertions(+), 4 deletions(-) create mode 100644 artifacts/vg_5k/natural_pipeline_orbit.json diff --git a/LAB_NOTES.md b/LAB_NOTES.md index d8d235f..8c0e9d5 100644 --- a/LAB_NOTES.md +++ b/LAB_NOTES.md @@ -549,10 +549,16 @@ statistics (0.19 for DINOv2 crop CLS against Qwen phrase states), so in-context features and corpus-specific distributional vectors are the better substrate. -Verdict: 0.656 against the 0.9 recovery needs; no recovery run. The -untried lever is observation orbits, which VG lacks by construction -- -one photograph per scene. Multi-view natural data is the next protocol -requirement. +Augmentation orbits were then tried and give nothing: four random resized +crops per image, segmented independently with fields averaged, move the +correlation from 0.6559 to 0.6559. Re-renders in the closed world +resample layout, which is nuisance by construction; random crops perturb +framing, which self-supervised patch features already absorb, so the +average has nothing to cancel. Multi-view observation means separate +photographs, not transformations of one. + +Verdict: 0.656 against the 0.9 recovery needs; no recovery run, and the +orbit lever is unavailable on this corpus. See `NATURAL_RESULTS.md`. diff --git a/NATURAL_RESULTS.md b/NATURAL_RESULTS.md index ef7339c..650c77e 100644 --- a/NATURAL_RESULTS.md +++ b/NATURAL_RESULTS.md @@ -86,6 +86,22 @@ PPMI vectors fitted to this corpus beat a general language model's states, echoing the earlier finding that what the model layer shares is the mixed state rather than the isolated encoding. +## Augmentation orbits are not a substitute for multiple photographs + +The closed world made two-sided multi-view observation a protocol +requirement, and Visual Genome carries one photograph per scene, so the +orbit was synthesised: four random resized crops per image, segmented +independently, their fields averaged. The correlation moves from 0.6559 +to 0.6559 -- no gain at four views. + +The mechanism is visible in what each orbit cancels. Re-renders in the +closed world resample layout, which is nuisance by construction, so +averaging over them removes exactly the modality-private variation. +Random crops perturb framing, and self-supervised patch features are +already close to invariant under it, so the average has nothing to +cancel. Multi-view observation, where it matters, means separate +photographs of the same scene rather than transformations of one. + ## Verdict The go/no-go statistic stands at 0.656 against the 0.9 that polynomial diff --git a/artifacts/vg_5k/natural_pipeline_orbit.json b/artifacts/vg_5k/natural_pipeline_orbit.json new file mode 100644 index 0000000..0365ac1 --- /dev/null +++ b/artifacts/vg_5k/natural_pipeline_orbit.json @@ -0,0 +1,8 @@ +{ + "protocol": "Continuous states, content projection fitted per modality on scenes outside the evaluated set. Hidden pairs are read only for the correlation.", + "samples": 256, + "views": 4, + "kept_directions": 8, + "field_correlation_at_truth": 0.6559307692721238, + "recovery_threshold": 0.9 +} \ No newline at end of file -- cgit v1.2.3