summaryrefslogtreecommitdiff
path: root/results
diff options
context:
space:
mode:
authorYurenHao0426 <Blackhao0426@gmail.com>2026-04-07 23:42:06 -0500
committerYurenHao0426 <Blackhao0426@gmail.com>2026-04-07 23:42:06 -0500
commit4cd716757b50a1f4217a3ffdf8ee624c270b7a23 (patch)
tree4754c2a6fc28f6a91f724498e2eaf7598a2f12fa /results
parentacc86add44e0cac8701307f936029770edd50891 (diff)
Add CNN third-architecture audit: BN, no terminal LN
5 methods × 3 seeds on the SmallCNN (3 conv + BN + 1 FC + head, no terminal LN) using existing checkpoints in results/cnn_baseline/. Key findings: BP CNN: 0.866 acc, max/block 1.3, trustworthy State Bridge CNN: 0.633 acc, max/block 2.4, trustworthy EP CNN: 0.512 acc, max/block 12, trustworthy DFA CNN: 0.566 acc, max/block 237, walked back via (a) Credit Bridge CNN: 0.325 acc, max/block 96, walked back via (a) CRITICAL: diagnostic (b) ||g_L|| floor NEVER fires on CNN for any method. The deepest BP grad is at ~1e-5 to 6e-1, all well above the 1e-7 floor. This is the cleanest confirmation that terminal LayerNorm is the structural cause of the catastrophic gradient collapse in (b). Without out_ln, the BP grad does NOT collapse to the floor, even on DFA. The scale pathology (a) still appears on DFA and CB, but the gradient collapse pathology (b) is specific to terminal-LN architectures. DFA CNN's accuracy (56.6%) is much higher than DFA ResMLP (30.8%) or DFA ViT (23.7%) — partially because the scale pathology is less catastrophic without the LN-driven gradient cancellation amplifying it. This is the cross-architecture mechanism story made concrete.
Diffstat (limited to 'results')
-rw-r--r--results/protocol_audit/audit_cnn_3seed.json287
1 files changed, 287 insertions, 0 deletions
diff --git a/results/protocol_audit/audit_cnn_3seed.json b/results/protocol_audit/audit_cnn_3seed.json
new file mode 100644
index 0000000..e31d90a
--- /dev/null
+++ b/results/protocol_audit/audit_cnn_3seed.json
@@ -0,0 +1,287 @@
+[
+ {
+ "method": "bp",
+ "seed": 42,
+ "acc": 0.8621,
+ "h_norms": [
+ 92.19713592529297,
+ 36.073795318603516,
+ 31.53321647644043,
+ 37.99840545654297
+ ],
+ "g_norms": [
+ 0.0003265859850216657,
+ 0.0005140144494362175,
+ 0.00032046454725787044,
+ 4.230972990626469e-05
+ ],
+ "max_per_block": 1.2050278944722594,
+ "verdict": "trustworthy"
+ },
+ {
+ "method": "dfa",
+ "seed": 42,
+ "acc": 0.5526,
+ "h_norms": [
+ 250.00730895996094,
+ 312.03765869140625,
+ 338.56951904296875,
+ 72491.6875
+ ],
+ "g_norms": [
+ 0.008991315960884094,
+ 0.004240375477820635,
+ 0.0019398012664169073,
+ 0.0007538454374298453
+ ],
+ "max_per_block": 214.1116769900361,
+ "verdict": "walk-back: (a)"
+ },
+ {
+ "method": "state_bridge",
+ "seed": 42,
+ "acc": 0.632,
+ "h_norms": [
+ 82.06472778320312,
+ 58.023536682128906,
+ 63.79484176635742,
+ 146.3084716796875
+ ],
+ "g_norms": [
+ 0.012001844123005867,
+ 0.0053405677899718285,
+ 0.002992230001837015,
+ 0.0019513164879754186
+ ],
+ "max_per_block": 2.2934216564958096,
+ "verdict": "trustworthy"
+ },
+ {
+ "method": "credit_bridge",
+ "seed": 42,
+ "acc": 0.3357,
+ "h_norms": [
+ 188.0859375,
+ 184.6356658935547,
+ 197.04025268554688,
+ 21333.876953125
+ ],
+ "g_norms": [
+ 0.011781061999499798,
+ 0.009520439431071281,
+ 0.007203294429928064,
+ 0.0029390468262135983
+ ],
+ "max_per_block": 108.27166866848961,
+ "verdict": "walk-back: (a)"
+ },
+ {
+ "method": "ep",
+ "seed": 42,
+ "acc": 0.5033,
+ "h_norms": [
+ 83.98963165283203,
+ 80.78277587890625,
+ 69.89965057373047,
+ 23.533870697021484
+ ],
+ "g_norms": [
+ 0.02113482914865017,
+ 0.01996646821498871,
+ 0.04826148599386215,
+ 0.6656972765922546
+ ],
+ "max_per_block": 0.9618184326943926,
+ "verdict": "trustworthy"
+ },
+ {
+ "method": "bp",
+ "seed": 123,
+ "acc": 0.8683,
+ "h_norms": [
+ 95.01183319091797,
+ 37.34379196166992,
+ 30.323930740356445,
+ 41.04403305053711
+ ],
+ "g_norms": [
+ 0.00020042041433043778,
+ 0.0002769956481643021,
+ 0.0002367425913689658,
+ 2.9566466764663346e-05
+ ],
+ "max_per_block": 1.3535195487013123,
+ "verdict": "trustworthy"
+ },
+ {
+ "method": "dfa",
+ "seed": 123,
+ "acc": 0.5501,
+ "h_norms": [
+ 259.4548034667969,
+ 386.1385498046875,
+ 345.1282958984375,
+ 81147.859375
+ ],
+ "g_norms": [
+ 0.0176572035998106,
+ 0.006263771094381809,
+ 0.004082133527845144,
+ 0.0012416314566507936
+ ],
+ "max_per_block": 235.12375061498798,
+ "verdict": "walk-back: (a)"
+ },
+ {
+ "method": "state_bridge",
+ "seed": 123,
+ "acc": 0.6277,
+ "h_norms": [
+ 78.61927032470703,
+ 61.77223587036133,
+ 62.21416473388672,
+ 168.63580322265625
+ ],
+ "g_norms": [
+ 0.01707093045115471,
+ 0.005183099303394556,
+ 0.0028397536370903254,
+ 0.001994567457586527
+ ],
+ "max_per_block": 2.7105692721902597,
+ "verdict": "trustworthy"
+ },
+ {
+ "method": "credit_bridge",
+ "seed": 123,
+ "acc": 0.3132,
+ "h_norms": [
+ 153.3919219970703,
+ 174.97061157226562,
+ 205.44915771484375,
+ 18492.830078125
+ ],
+ "g_norms": [
+ 0.011560788378119469,
+ 0.00875047780573368,
+ 0.007237838581204414,
+ 0.0031556652393192053
+ ],
+ "max_per_block": 90.01171036092735,
+ "verdict": "walk-back: (a)"
+ },
+ {
+ "method": "ep",
+ "seed": 123,
+ "acc": 0.4897,
+ "h_norms": [
+ 86.96959686279297,
+ 82.86459350585938,
+ 57.40688705444336,
+ 1.998042345046997
+ ],
+ "g_norms": [
+ 0.01942148432135582,
+ 0.0226480383425951,
+ 0.06046672910451889,
+ 1.157748818397522
+ ],
+ "max_per_block": 0.9527995586387525,
+ "verdict": "trustworthy"
+ },
+ {
+ "method": "bp",
+ "seed": 456,
+ "acc": 0.8681,
+ "h_norms": [
+ 96.83692169189453,
+ 37.44154739379883,
+ 31.123756408691406,
+ 42.27854919433594
+ ],
+ "g_norms": [
+ 0.00014156661927700043,
+ 0.00021504472533706576,
+ 0.00015155959408730268,
+ 1.7612564988667145e-05
+ ],
+ "max_per_block": 1.3584012366363825,
+ "verdict": "trustworthy"
+ },
+ {
+ "method": "dfa",
+ "seed": 456,
+ "acc": 0.5954,
+ "h_norms": [
+ 206.5431671142578,
+ 266.57421875,
+ 254.53468322753906,
+ 66974.734375
+ ],
+ "g_norms": [
+ 0.00448678620159626,
+ 0.002167485887184739,
+ 0.0012352537596598268,
+ 0.0004585048300214112
+ ],
+ "max_per_block": 263.12616232000306,
+ "verdict": "walk-back: (a)"
+ },
+ {
+ "method": "state_bridge",
+ "seed": 456,
+ "acc": 0.6396,
+ "h_norms": [
+ 71.60630798339844,
+ 56.15557098388672,
+ 63.141014099121094,
+ 137.78231811523438
+ ],
+ "g_norms": [
+ 0.014506030827760696,
+ 0.005259184632450342,
+ 0.0027562177274376154,
+ 0.001790599781088531
+ ],
+ "max_per_block": 2.1821366045679693,
+ "verdict": "trustworthy"
+ },
+ {
+ "method": "credit_bridge",
+ "seed": 456,
+ "acc": 0.3251,
+ "h_norms": [
+ 169.41067504882812,
+ 151.55250549316406,
+ 177.73605346679688,
+ 16139.1640625
+ ],
+ "g_norms": [
+ 0.009933868423104286,
+ 0.00681547075510025,
+ 0.004268169403076172,
+ 0.0029866439290344715
+ ],
+ "max_per_block": 90.80410950789441,
+ "verdict": "walk-back: (a)"
+ },
+ {
+ "method": "ep",
+ "seed": 456,
+ "acc": 0.5432,
+ "h_norms": [
+ 84.82034301757812,
+ 85.0868911743164,
+ 163.61471557617188,
+ 5375.328125
+ ],
+ "g_norms": [
+ 0.028629517182707787,
+ 0.01810075342655182,
+ 0.0257416944950819,
+ 0.3386228382587433
+ ],
+ "max_per_block": 32.85357375142385,
+ "verdict": "trustworthy"
+ }
+] \ No newline at end of file