Hugging Face's logo

Datasets:
SabaPivot
/
icml26-repro-nonlinear-autoencoder-pca

Modalities:
Tabular
Formats:
csv
Size:
< 1K
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community
Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
d
int64
64
256
seed
int64
0
3
epoch
int64
0
499
loss
float64
31.5
128
u_overlap
float64
0
0.83
v_overlap
float64
0
0.3
64
0
0
31.929258
0.047502
0.091713
64
0
10
31.713882
0.045371
0.049313
64
0
20
31.595713
0.040425
0.015587
64
0
30
31.535269
0.023717
0.045014
64
0
40
31.51156
0.106345
0.08266
64
0
50
31.50625
0.174236
0.074609
64
0
60
31.501215
0.245393
0.060381
64
0
70
31.497576
0.324335
0.038375
64
0
80
31.495436
0.400851
0.019092
64
0
90
31.493704
0.464785
0.007685
64
0
100
31.491913
0.525574
0.027311
64
0
110
31.489311
0.593012
0.050343
64
0
120
31.485413
0.66323
0.079978
64
0
130
31.480352
0.727728
0.110338
64
0
140
31.475348
0.783812
0.135649
64
0
150
31.473814
0.814703
0.153741
64
0
160
31.473692
0.825914
0.162526
64
0
170
31.473715
0.826492
0.165172
64
0
180
31.47368
0.824494
0.163224
64
0
190
31.473656
0.822776
0.161333
64
0
200
31.473648
0.821527
0.15987
64
0
210
31.473646
0.821054
0.159112
64
0
220
31.473648
0.820885
0.15907
64
0
230
31.473646
0.821033
0.159235
64
0
240
31.473648
0.821222
0.159457
64
0
250
31.473646
0.821331
0.159555
64
0
260
31.473646
0.821359
0.159585
64
0
270
31.47365
0.821339
0.159568
64
0
280
31.47365
0.821316
0.159547
64
0
290
31.473646
0.821308
0.159533
64
0
300
31.473648
0.821302
0.159527
64
0
310
31.47365
0.821303
0.159528
64
0
320
31.473646
0.821306
0.159532
64
0
330
31.473646
0.821308
0.159534
64
0
340
31.473646
0.821308
0.159534
64
0
350
31.473646
0.821308
0.159534
64
0
360
31.473648
0.821308
0.159534
64
0
370
31.473646
0.821308
0.159533
64
0
380
31.473646
0.821308
0.159533
64
0
390
31.473646
0.821308
0.159533
64
0
400
31.473646
0.821308
0.159533
64
0
410
31.473646
0.821308
0.159533
64
0
420
31.473646
0.821308
0.159533
64
0
430
31.473646
0.821308
0.159533
64
0
440
31.473646
0.821308
0.159533
64
0
450
31.473646
0.821308
0.159533
64
0
460
31.473646
0.821308
0.159533
64
0
470
31.473646
0.821308
0.159533
64
0
480
31.473646
0.821308
0.159533
64
0
490
31.473646
0.821308
0.159533
64
0
499
31.473646
0.821308
0.159533
64
1
0
32.295055
0.323627
0.11058
64
1
10
32.108307
0.507785
0.106591
64
1
20
31.96829
0.611458
0.079535
64
1
30
31.911886
0.635049
0.054185
64
1
40
31.903067
0.598505
0.027125
64
1
50
31.898493
0.572866
0.024355
64
1
60
31.895914
0.570892
0.032344
64
1
70
31.895124
0.563035
0.031239
64
1
80
31.894747
0.548382
0.024446
64
1
90
31.894503
0.533167
0.019497
64
1
100
31.894297
0.519281
0.012279
64
1
110
31.894144
0.505759
0.005631
64
1
120
31.894073
0.495204
0.000069
64
1
130
31.894035
0.487376
0.003859
64
1
140
31.894022
0.482266
0.00577
64
1
150
31.894018
0.478992
0.00743
64
1
160
31.894012
0.476435
0.009016
64
1
170
31.894011
0.474476
0.010166
64
1
180
31.894007
0.472778
0.011061
64
1
190
31.894007
0.471303
0.011831
64
1
200
31.894007
0.470001
0.012544
64
1
210
31.894007
0.468828
0.013156
64
1
220
31.894005
0.467819
0.013691
64
1
230
31.894005
0.466954
0.014169
64
1
240
31.894005
0.466265
0.014542
64
1
250
31.894005
0.465747
0.014826
64
1
260
31.894005
0.465367
0.015038
64
1
270
31.894003
0.465101
0.015184
64
1
280
31.894005
0.46492
0.015281
64
1
290
31.894005
0.464801
0.015346
64
1
300
31.894005
0.464727
0.015387
64
1
310
31.894003
0.464683
0.015411
64
1
320
31.894005
0.464657
0.015425
64
1
330
31.894007
0.464643
0.015433
64
1
340
31.894005
0.464636
0.015437
64
1
350
31.894005
0.464632
0.015439
64
1
360
31.894005
0.464631
0.015439
64
1
370
31.894005
0.464631
0.015439
64
1
380
31.894005
0.464631
0.015439
64
1
390
31.894005
0.464631
0.015439
64
1
400
31.894005
0.464631
0.015439
64
1
410
31.894003
0.464631
0.015439
64
1
420
31.894003
0.464632
0.015439
64
1
430
31.894003
0.464632
0.015439
64
1
440
31.894005
0.464632
0.015439
64
1
450
31.894003
0.464632
0.015439
64
1
460
31.894003
0.464632
0.015439
64
1
470
31.894005
0.464632
0.015439
64
1
480
31.894005
0.464632
0.015439
End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Nonlinear autoencoder / PCA reproduction

This bundle reproduces the paper with the authors' pinned release (SPOC-group/advantage_nonlinearity@4378017) plus an independent direct population-gradient-flow audit.

  • run_official_amp_campaign.py: 96 finite-dimensional runs of the released two-spike AMP at d=256/512 and eight sample ratios.
  • run_official_ae_campaign.py: 48 full-batch Adam runs of the released tied, one-neuron ReLU/ELU autoencoder at d=512/1024, with exact PCA and 20,000 held-out examples per run.
  • run_population_gradient_flow.py: deterministic quadrature of the actual population loss and RK4 integration of its spherical projected gradient.
  • summarize_official_campaign.py: aggregate metrics and the summary figure.

The original smaller independent check remains in reproduce.py for audit history, but the claim verdicts use the official-code campaign above.

Rerun 

uv run --with-requirements requirements.txt python run_official_amp_campaign.py
uv run --with-requirements requirements.txt python run_official_ae_campaign.py
uv run --with-requirements requirements.txt python run_population_gradient_flow.py
uv run --with-requirements requirements.txt python summarize_official_campaign.py

The AE campaign uses CUDA when available and took about 10.1 minutes on one H100. AMP and population-flow checks run on CPU. All seeds and exact grids are inside the scripts and every raw run is retained under results/.
Downloads last month
7
Number of rows:
612
Total file size:
1.83 MB