diff options
| author | YurenHao0426 <Blackhao0426@gmail.com> | 2026-06-14 20:32:31 -0500 |
|---|---|---|
| committer | YurenHao0426 <Blackhao0426@gmail.com> | 2026-06-14 20:32:31 -0500 |
| commit | 1118b7457c261de36ead6103503c00c321c75f9b (patch) | |
| tree | 7ea76b32f070cb58458caaa2897a5d8133561f48 /results/depth_ladder/ladder_d512_L2_cifar10.json | |
| parent | aa73718eb6427d7da3b9cb416275802d90c4b2ed (diff) | |
Appendix experiment triangulating the depth-utility diagnostic (D3) by varying
the number of trainable residual blocks k (last-k trainable, first L-k frozen at
init; embed/LN/head always trained).
- d=256 L=4 and d=512 L=2, 3 seeds, recipe identical to the main audit.
- BP climbs monotonically (+22-23pp); DFA peaks at the frozen baseline (k=0) and
declines once any deep block is trained; FA shows partial/no net depth utility.
- Cross-checks reproduce existing anchors (BP 0.617, DFA 0.301, FA 0.402, frozen 0.349).
- frozen_init_identity_check quantifies frozen stack as a near-norm-preserving
random feature map (per-block ||f||/||h||~0.10, stack cos 0.981), explaining the
above-chance k=0 rung.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Diffstat (limited to 'results/depth_ladder/ladder_d512_L2_cifar10.json')
| -rw-r--r-- | results/depth_ladder/ladder_d512_L2_cifar10.json | 1378 |
1 files changed, 1378 insertions, 0 deletions
diff --git a/results/depth_ladder/ladder_d512_L2_cifar10.json b/results/depth_ladder/ladder_d512_L2_cifar10.json new file mode 100644 index 0000000..4a3feff --- /dev/null +++ b/results/depth_ladder/ladder_d512_L2_cifar10.json @@ -0,0 +1,1378 @@ +{ + "config": { + "d_hidden": 512, + "num_blocks": 2, + "dataset": "cifar10", + "methods": [ + "bp", + "fa", + "dfa" + ], + "k_values": [ + 0, + 1, + 2 + ], + "seeds": [ + 42, + 123, + 456 + ], + "epochs": 100, + "lr": 0.001, + "wd": 0.01, + "batch_size": 128, + "gpu": 0, + "output_dir": "results/depth_ladder", + "num_classes": 10 + }, + "results": { + "bp": { + "0": { + "42": { + "final_acc": 0.3891, + "curve": [ + [ + 1, + 0.3462 + ], + [ + 10, + 0.3633 + ], + [ + 20, + 0.3635 + ], + [ + 30, + 0.3543 + ], + [ + 40, + 0.3673 + ], + [ + 50, + 0.3633 + ], + [ + 60, + 0.3695 + ], + [ + 70, + 0.3753 + ], + [ + 80, + 0.3858 + ], + [ + 90, + 0.3887 + ], + [ + 100, + 0.3891 + ] + ] + }, + "123": { + "final_acc": 0.3846, + "curve": [ + [ + 1, + 0.3497 + ], + [ + 10, + 0.3704 + ], + [ + 20, + 0.3698 + ], + [ + 30, + 0.354 + ], + [ + 40, + 0.3505 + ], + [ + 50, + 0.3634 + ], + [ + 60, + 0.3675 + ], + [ + 70, + 0.3739 + ], + [ + 80, + 0.3823 + ], + [ + 90, + 0.3845 + ], + [ + 100, + 0.3846 + ] + ] + }, + "456": { + "final_acc": 0.3838, + "curve": [ + [ + 1, + 0.3409 + ], + [ + 10, + 0.3578 + ], + [ + 20, + 0.3767 + ], + [ + 30, + 0.3607 + ], + [ + 40, + 0.3551 + ], + [ + 50, + 0.3632 + ], + [ + 60, + 0.3722 + ], + [ + 70, + 0.3704 + ], + [ + 80, + 0.3784 + ], + [ + 90, + 0.3834 + ], + [ + 100, + 0.3838 + ] + ] + } + }, + "1": { + "42": { + "final_acc": 0.5856, + "curve": [ + [ + 1, + 0.3667 + ], + [ + 10, + 0.4836 + ], + [ + 20, + 0.5197 + ], + [ + 30, + 0.5367 + ], + [ + 40, + 0.5444 + ], + [ + 50, + 0.5629 + ], + [ + 60, + 0.5691 + ], + [ + 70, + 0.5779 + ], + [ + 80, + 0.5808 + ], + [ + 90, + 0.5849 + ], + [ + 100, + 0.5856 + ] + ] + }, + "123": { + "final_acc": 0.5819, + "curve": [ + [ + 1, + 0.3632 + ], + [ + 10, + 0.4865 + ], + [ + 20, + 0.5175 + ], + [ + 30, + 0.536 + ], + [ + 40, + 0.5466 + ], + [ + 50, + 0.5606 + ], + [ + 60, + 0.5716 + ], + [ + 70, + 0.5749 + ], + [ + 80, + 0.5806 + ], + [ + 90, + 0.5817 + ], + [ + 100, + 0.5819 + ] + ] + }, + "456": { + "final_acc": 0.5809, + "curve": [ + [ + 1, + 0.3696 + ], + [ + 10, + 0.4737 + ], + [ + 20, + 0.5199 + ], + [ + 30, + 0.5317 + ], + [ + 40, + 0.5498 + ], + [ + 50, + 0.561 + ], + [ + 60, + 0.5675 + ], + [ + 70, + 0.5767 + ], + [ + 80, + 0.5785 + ], + [ + 90, + 0.5802 + ], + [ + 100, + 0.5809 + ] + ] + } + }, + "2": { + "42": { + "final_acc": 0.6039, + "curve": [ + [ + 1, + 0.379 + ], + [ + 10, + 0.5174 + ], + [ + 20, + 0.5471 + ], + [ + 30, + 0.5712 + ], + [ + 40, + 0.5906 + ], + [ + 50, + 0.5969 + ], + [ + 60, + 0.5977 + ], + [ + 70, + 0.5992 + ], + [ + 80, + 0.6072 + ], + [ + 90, + 0.6037 + ], + [ + 100, + 0.6039 + ] + ] + }, + "123": { + "final_acc": 0.602, + "curve": [ + [ + 1, + 0.3732 + ], + [ + 10, + 0.5161 + ], + [ + 20, + 0.5554 + ], + [ + 30, + 0.5756 + ], + [ + 40, + 0.5811 + ], + [ + 50, + 0.5928 + ], + [ + 60, + 0.5965 + ], + [ + 70, + 0.6016 + ], + [ + 80, + 0.6027 + ], + [ + 90, + 0.6007 + ], + [ + 100, + 0.602 + ] + ] + }, + "456": { + "final_acc": 0.6045, + "curve": [ + [ + 1, + 0.3768 + ], + [ + 10, + 0.5097 + ], + [ + 20, + 0.5499 + ], + [ + 30, + 0.5773 + ], + [ + 40, + 0.5858 + ], + [ + 50, + 0.5845 + ], + [ + 60, + 0.5934 + ], + [ + 70, + 0.5985 + ], + [ + 80, + 0.6011 + ], + [ + 90, + 0.602 + ], + [ + 100, + 0.6045 + ] + ] + } + } + }, + "fa": { + "0": { + "42": { + "final_acc": 0.3585, + "curve": [ + [ + 1, + 0.3288 + ], + [ + 10, + 0.3359 + ], + [ + 20, + 0.3336 + ], + [ + 30, + 0.3328 + ], + [ + 40, + 0.3418 + ], + [ + 50, + 0.3504 + ], + [ + 60, + 0.3564 + ], + [ + 70, + 0.3567 + ], + [ + 80, + 0.3543 + ], + [ + 90, + 0.3574 + ], + [ + 100, + 0.3585 + ] + ] + }, + "123": { + "final_acc": 0.3584, + "curve": [ + [ + 1, + 0.3125 + ], + [ + 10, + 0.3374 + ], + [ + 20, + 0.3364 + ], + [ + 30, + 0.3453 + ], + [ + 40, + 0.3437 + ], + [ + 50, + 0.3522 + ], + [ + 60, + 0.3587 + ], + [ + 70, + 0.355 + ], + [ + 80, + 0.3551 + ], + [ + 90, + 0.3558 + ], + [ + 100, + 0.3584 + ] + ] + }, + "456": { + "final_acc": 0.359, + "curve": [ + [ + 1, + 0.318 + ], + [ + 10, + 0.3311 + ], + [ + 20, + 0.3344 + ], + [ + 30, + 0.3533 + ], + [ + 40, + 0.3476 + ], + [ + 50, + 0.3523 + ], + [ + 60, + 0.3455 + ], + [ + 70, + 0.3569 + ], + [ + 80, + 0.3562 + ], + [ + 90, + 0.3583 + ], + [ + 100, + 0.359 + ] + ] + } + }, + "1": { + "42": { + "final_acc": 0.4083, + "curve": [ + [ + 1, + 0.3235 + ], + [ + 10, + 0.373 + ], + [ + 20, + 0.3734 + ], + [ + 30, + 0.3829 + ], + [ + 40, + 0.3916 + ], + [ + 50, + 0.4008 + ], + [ + 60, + 0.4012 + ], + [ + 70, + 0.4015 + ], + [ + 80, + 0.4042 + ], + [ + 90, + 0.4082 + ], + [ + 100, + 0.4083 + ] + ] + }, + "123": { + "final_acc": 0.4134, + "curve": [ + [ + 1, + 0.293 + ], + [ + 10, + 0.3662 + ], + [ + 20, + 0.3905 + ], + [ + 30, + 0.4027 + ], + [ + 40, + 0.3948 + ], + [ + 50, + 0.4048 + ], + [ + 60, + 0.4067 + ], + [ + 70, + 0.4094 + ], + [ + 80, + 0.4115 + ], + [ + 90, + 0.4103 + ], + [ + 100, + 0.4134 + ] + ] + }, + "456": { + "final_acc": 0.4155, + "curve": [ + [ + 1, + 0.3098 + ], + [ + 10, + 0.3561 + ], + [ + 20, + 0.386 + ], + [ + 30, + 0.3957 + ], + [ + 40, + 0.3907 + ], + [ + 50, + 0.4032 + ], + [ + 60, + 0.4017 + ], + [ + 70, + 0.4125 + ], + [ + 80, + 0.4123 + ], + [ + 90, + 0.4164 + ], + [ + 100, + 0.4155 + ] + ] + } + }, + "2": { + "42": { + "final_acc": 0.3582, + "curve": [ + [ + 1, + 0.3028 + ], + [ + 10, + 0.3585 + ], + [ + 20, + 0.3523 + ], + [ + 30, + 0.3315 + ], + [ + 40, + 0.3191 + ], + [ + 50, + 0.3397 + ], + [ + 60, + 0.3566 + ], + [ + 70, + 0.3527 + ], + [ + 80, + 0.3554 + ], + [ + 90, + 0.3593 + ], + [ + 100, + 0.3582 + ] + ] + }, + "123": { + "final_acc": 0.3621, + "curve": [ + [ + 1, + 0.2794 + ], + [ + 10, + 0.3627 + ], + [ + 20, + 0.36 + ], + [ + 30, + 0.375 + ], + [ + 40, + 0.3482 + ], + [ + 50, + 0.3679 + ], + [ + 60, + 0.363 + ], + [ + 70, + 0.3643 + ], + [ + 80, + 0.3636 + ], + [ + 90, + 0.3618 + ], + [ + 100, + 0.3621 + ] + ] + }, + "456": { + "final_acc": 0.3642, + "curve": [ + [ + 1, + 0.3005 + ], + [ + 10, + 0.3573 + ], + [ + 20, + 0.3624 + ], + [ + 30, + 0.3706 + ], + [ + 40, + 0.3529 + ], + [ + 50, + 0.3648 + ], + [ + 60, + 0.3581 + ], + [ + 70, + 0.3645 + ], + [ + 80, + 0.3652 + ], + [ + 90, + 0.3632 + ], + [ + 100, + 0.3642 + ] + ] + } + } + }, + "dfa": { + "0": { + "42": { + "final_acc": 0.3432, + "curve": [ + [ + 1, + 0.3196 + ], + [ + 10, + 0.3187 + ], + [ + 20, + 0.3369 + ], + [ + 30, + 0.3221 + ], + [ + 40, + 0.3386 + ], + [ + 50, + 0.3401 + ], + [ + 60, + 0.3473 + ], + [ + 70, + 0.3472 + ], + [ + 80, + 0.3426 + ], + [ + 90, + 0.3445 + ], + [ + 100, + 0.3432 + ] + ] + }, + "123": { + "final_acc": 0.3508, + "curve": [ + [ + 1, + 0.3089 + ], + [ + 10, + 0.318 + ], + [ + 20, + 0.3301 + ], + [ + 30, + 0.3434 + ], + [ + 40, + 0.3386 + ], + [ + 50, + 0.3343 + ], + [ + 60, + 0.3489 + ], + [ + 70, + 0.3458 + ], + [ + 80, + 0.3499 + ], + [ + 90, + 0.3508 + ], + [ + 100, + 0.3508 + ] + ] + }, + "456": { + "final_acc": 0.3521, + "curve": [ + [ + 1, + 0.3238 + ], + [ + 10, + 0.3327 + ], + [ + 20, + 0.3395 + ], + [ + 30, + 0.3457 + ], + [ + 40, + 0.3367 + ], + [ + 50, + 0.3496 + ], + [ + 60, + 0.3453 + ], + [ + 70, + 0.3487 + ], + [ + 80, + 0.3491 + ], + [ + 90, + 0.3498 + ], + [ + 100, + 0.3521 + ] + ] + } + }, + "1": { + "42": { + "final_acc": 0.2384, + "curve": [ + [ + 1, + 0.2687 + ], + [ + 10, + 0.2106 + ], + [ + 20, + 0.2293 + ], + [ + 30, + 0.2297 + ], + [ + 40, + 0.2241 + ], + [ + 50, + 0.2318 + ], + [ + 60, + 0.2417 + ], + [ + 70, + 0.2458 + ], + [ + 80, + 0.2463 + ], + [ + 90, + 0.2438 + ], + [ + 100, + 0.2384 + ] + ] + }, + "123": { + "final_acc": 0.2097, + "curve": [ + [ + 1, + 0.1958 + ], + [ + 10, + 0.1777 + ], + [ + 20, + 0.222 + ], + [ + 30, + 0.1852 + ], + [ + 40, + 0.2165 + ], + [ + 50, + 0.2095 + ], + [ + 60, + 0.1995 + ], + [ + 70, + 0.2038 + ], + [ + 80, + 0.2068 + ], + [ + 90, + 0.2173 + ], + [ + 100, + 0.2097 + ] + ] + }, + "456": { + "final_acc": 0.2295, + "curve": [ + [ + 1, + 0.2118 + ], + [ + 10, + 0.2074 + ], + [ + 20, + 0.1777 + ], + [ + 30, + 0.2043 + ], + [ + 40, + 0.201 + ], + [ + 50, + 0.2087 + ], + [ + 60, + 0.2073 + ], + [ + 70, + 0.2126 + ], + [ + 80, + 0.2202 + ], + [ + 90, + 0.2355 + ], + [ + 100, + 0.2295 + ] + ] + } + }, + "2": { + "42": { + "final_acc": 0.3069, + "curve": [ + [ + 1, + 0.2769 + ], + [ + 10, + 0.2705 + ], + [ + 20, + 0.3 + ], + [ + 30, + 0.2988 + ], + [ + 40, + 0.308 + ], + [ + 50, + 0.2941 + ], + [ + 60, + 0.3025 + ], + [ + 70, + 0.3075 + ], + [ + 80, + 0.307 + ], + [ + 90, + 0.3063 + ], + [ + 100, + 0.3069 + ] + ] + }, + "123": { + "final_acc": 0.3025, + "curve": [ + [ + 1, + 0.2582 + ], + [ + 10, + 0.2772 + ], + [ + 20, + 0.2904 + ], + [ + 30, + 0.3072 + ], + [ + 40, + 0.2898 + ], + [ + 50, + 0.2938 + ], + [ + 60, + 0.2892 + ], + [ + 70, + 0.2974 + ], + [ + 80, + 0.297 + ], + [ + 90, + 0.3035 + ], + [ + 100, + 0.3025 + ] + ] + }, + "456": { + "final_acc": 0.2963, + "curve": [ + [ + 1, + 0.2794 + ], + [ + 10, + 0.2888 + ], + [ + 20, + 0.2884 + ], + [ + 30, + 0.2901 + ], + [ + 40, + 0.2784 + ], + [ + 50, + 0.2817 + ], + [ + 60, + 0.2983 + ], + [ + 70, + 0.292 + ], + [ + 80, + 0.2904 + ], + [ + 90, + 0.2999 + ], + [ + 100, + 0.2963 + ] + ] + } + } + } + } +}
\ No newline at end of file |
