geolip-conduit-experiments / cell_4_proper_experiment_3.txt

Create cell_4_proper_experiment_3.txt

1b9a2c2 verified 11 days ago

10.6 kB

	Loading Freckles v40 + CIFAR-10...

	======================================================================
	1. FULL ROUND-TRIP — Per-patch reconstruction error
	======================================================================

	Collecting per-patch reconstruction errors...


	Reconstructing: 0%\| \| 0/157 [00:00<?, ?it/s]
	Reconstructing: 3%\|▎ \| 4/157 [00:00<00:03, 38.85it/s]
	Reconstructing: 8%\|▊ \| 13/157 [00:00<00:02, 68.11it/s]
	Reconstructing: 14%\|█▍ \| 22/157 [00:00<00:01, 77.91it/s]
	Reconstructing: 20%\|█▉ \| 31/157 [00:00<00:01, 81.98it/s]
	Reconstructing: 26%\|██▌ \| 41/157 [00:00<00:01, 85.54it/s]
	Reconstructing: 32%\|███▏ \| 51/157 [00:00<00:01, 87.75it/s]
	Reconstructing: 39%\|███▉ \| 61/157 [00:00<00:01, 89.17it/s]
	Reconstructing: 45%\|████▌ \| 71/157 [00:00<00:00, 89.94it/s]
	Reconstructing: 52%\|█████▏ \| 81/157 [00:00<00:00, 90.52it/s]
	Reconstructing: 58%\|█████▊ \| 91/157 [00:01<00:00, 90.89it/s]
	Reconstructing: 64%\|██████▍ \| 101/157 [00:01<00:00, 91.03it/s]
	Reconstructing: 71%\|███████ \| 111/157 [00:01<00:00, 91.05it/s]
	Reconstructing: 77%\|███████▋ \| 121/157 [00:01<00:00, 91.14it/s]
	Reconstructing: 83%\|████████▎ \| 131/157 [00:01<00:00, 91.22it/s]
	Reconstructing: 90%\|████████▉ \| 141/157 [00:01<00:00, 91.26it/s]
	Reconstructing: 100%\|██████████\| 157/157 [00:01<00:00, 86.32it/s]
	Collected 10000 images, 2000 individual maps

	======================================================================
	1a. SPATIAL STRUCTURE — Does recon error vary across patches?
	======================================================================

	Per-image spatial CV of reconstruction error:
	Mean CV: 0.3972
	Median CV: 0.3989
	Min CV: 0.0865
	Max CV: 0.7128
	VERDICT: HAS SPATIAL STRUCTURE

	======================================================================
	1b. PER-CLASS RECONSTRUCTION ERROR
	======================================================================

	Class Mean MSE Std MSE Max patch
	------------------------------------------
	airplane 0.000000 0.000000 0.000000
	auto 0.000000 0.000000 0.000000
	bird 0.000000 0.000000 0.000000
	cat 0.000000 0.000000 0.000000
	deer 0.000000 0.000000 0.000000
	dog 0.000000 0.000000 0.000000
	frog 0.000000 0.000000 0.000000
	horse 0.000000 0.000000 0.000000
	ship 0.000000 0.000000 0.000000
	truck 0.000000 0.000000 0.000000

	Mean inter-class cosine similarity: 0.996998
	Min inter-class cosine similarity: 0.991408
	VERDICT: SIMILAR PATTERNS

	======================================================================
	2. CENTER vs EDGE — Where does reconstruction fail?
	======================================================================

	Class Center Edge Corner E/C ratio
	------------------------------------------------
	airplane 0.000000 0.000000 0.000000 0.9007
	auto 0.000000 0.000000 0.000000 0.9717
	bird 0.000000 0.000000 0.000000 0.9379
	cat 0.000000 0.000000 0.000000 0.9448
	deer 0.000000 0.000000 0.000000 0.9685
	dog 0.000000 0.000000 0.000000 1.0470
	frog 0.000000 0.000000 0.000000 0.9538
	horse 0.000000 0.000000 0.000000 0.9497
	ship 0.000000 0.000000 0.000000 1.0124
	truck 0.000000 0.000000 0.000000 0.9136

	======================================================================
	3. PER-MODE RECONSTRUCTION — Ablating SVD modes
	======================================================================

	Reconstructing with individual modes...

	Per-mode energy fraction (how much each mode contributes):

	Class Mode0 Mode1 Mode2 Mode3 FullMSE
	--------------------------------------------------
	airplane 0.4242 0.3352 0.1705 0.0701 0.000000
	auto 0.4234 0.3359 0.1704 0.0703 0.000000
	bird 0.4237 0.3361 0.1700 0.0703 0.000000
	cat 0.4232 0.3363 0.1701 0.0704 0.000000
	deer 0.4236 0.3363 0.1700 0.0701 0.000000
	dog 0.4238 0.3358 0.1703 0.0701 0.000000
	frog 0.4229 0.3367 0.1698 0.0706 0.000000
	horse 0.4237 0.3358 0.1703 0.0702 0.000000
	ship 0.4243 0.3353 0.1706 0.0699 0.000000
	truck 0.4237 0.3358 0.1704 0.0701 0.000000

	======================================================================
	4. LINEAR PROBE — Reconstruction error maps as features
	======================================================================

	Ridge probe comparison:

	Recon error spatial map dims= 256 train=51.6% test=19.0%
	Class Acc
	------------------
	airplane 25.9% █████
	auto 19.4% ███
	bird 9.1% █
	cat 12.2% ██
	deer 7.9% █
	dog 31.4% ██████
	frog 22.7% ████
	horse 14.3% ██
	ship 36.6% ███████
	truck 15.6% ███

	======================================================================
	5. FULL CONDUIT — Release error + eigenvalues + friction
	======================================================================

	Full conduit: 0%\| \| 0/157 [00:00<?, ?it/s]
	Full conduit: 2%\|▏ \| 3/157 [00:00<00:06, 25.60it/s]
	Full conduit: 6%\|▌ \| 9/157 [00:00<00:03, 42.69it/s]
	Full conduit: 10%\|▉ \| 15/157 [00:00<00:02, 48.80it/s]
	Full conduit: 13%\|█▎ \| 21/157 [00:00<00:02, 51.60it/s]
	Full conduit: 20%\|██ \| 32/157 [00:00<00:02, 47.18it/s]

	Comparative linear probes:

	Release error only dims= 256 train=51.6% test=16.0%
	Class Acc
	------------------
	airplane 27.8% █████
	auto 18.2% ███
	bird 14.6% ██
	cat 5.4% █
	deer 8.3% █
	dog 21.6% ████
	frog 18.9% ███
	horse 17.1% ███
	ship 11.6% ██
	truck 20.0% ████

	Eigenvalues (S) only dims= 1024 train=91.6% test=20.7%
	Class Acc
	------------------
	airplane 16.7% ███
	auto 24.2% ████
	bird 18.8% ███
	cat 21.6% ████
	deer 14.6% ██
	dog 16.2% ███
	frog 24.3% ████
	horse 17.1% ███
	ship 32.6% ██████
	truck 22.5% ████

	Friction only dims= 1024 train=92.4% test=22.5%
	Class Acc
	------------------
	airplane 19.4% ███
	auto 21.2% ████
	bird 18.8% ███
	cat 21.6% ████
	deer 14.6% ██
	dog 16.2% ███
	frog 27.0% █████
	horse 22.0% ████
	ship 34.9% ██████
	truck 30.0% ██████

	Combinations:

	Release + Eigenvalues dims= 1280 train=98.4% test=18.5%
	Class Acc
	------------------
	airplane 2.8%
	auto 21.2% ████
	bird 18.8% ███
	cat 2.7%
	deer 18.8% ███
	dog 16.2% ███
	frog 21.6% ████
	horse 26.8% █████
	ship 25.6% █████
	truck 27.5% █████
	Release + Friction dims= 1280 train=98.3% test=15.3%
	Class Acc
	------------------
	airplane 8.3% █
	auto 18.2% ███
	bird 16.7% ███
	cat 5.4% █
	deer 12.5% ██
	dog 16.2% ███
	frog 13.5% ██
	horse 19.5% ███
	ship 23.3% ████
	truck 17.5% ███
	Release + Eigenvalues + Friction dims= 2304 train=99.9% test=17.0%
	Class Acc
	------------------
	airplane 13.9% ██
	auto 3.0%
	bird 16.7% ███
	cat 10.8% ██
	deer 12.5% ██
	dog 21.6% ████
	frog 27.0% █████
	horse 22.0% ████
	ship 23.3% ████
	truck 17.5% ███

	======================================================================
	6. HIGH-ERROR PATCHES — Where does reconstruction fail?
	======================================================================

	Top error positions per class (patch coordinates):
	Class Top 3 positions (row, col) Error ratio
	----------------------------------------------------------------
	airplane (10,8), (9,8), (10,7) 1.19x
	auto (11,14), (11,7), (10,1) 1.12x
	bird (8,9), (7,9), (10,8) 1.11x
	cat (5,8), (5,7), (6,5) 1.08x
	deer (5,7), (5,5), (7,6) 1.06x
	dog (15,15), (14,15), (9,15) 1.07x
	frog (5,4), (4,7), (4,6) 1.07x
	horse (10,9), (9,7), (9,9) 1.11x
	ship (14,13), (14,12), (15,9) 1.20x
	truck (10,14), (10,13), (10,1) 1.16x

	Overall error map:
	Mean: 0.000000
	Std: 0.000000
	Hot patches (>2σ): 0/256

	======================================================================
	THEOREM 3: RELEASE FIDELITY — SUMMARY
	======================================================================

	SPATIAL STRUCTURE:
	Recon error spatial CV: 0.3972
	(Friction spatial CV was: 0.0137)

	CLASSIFICATION (ridge probe, test accuracy):
	Chance: 10.0%
	Friction maps: 24.3% (from Cell 3)
	Eigenvalue (S) maps: 21.0% (from Cell 3)
	Release error maps: 19.0%
	Release + Eigenvalues: 18.5%
	Release + Friction: 15.3%
	FULL CONDUIT (all three): 17.0%

	THE QUESTION ANSWERED:
	Does the release signal carry class-discriminative information
	that eigenvalues and friction do not?
	Lift from release over eigenvalues: -4.7pp
	Lift from full conduit over eigenvalues: -3.7pp