1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
|
=== CIFAR-100 PROTOCOL VALIDATION ===
Start: Wed Apr 29 09:24:37 PM CDT 2026
--- BP + FA + DFA on CIFAR-100 ---
Using device: cuda:0
============================================================
Seed 42
============================================================
--- BP ---
[BP] Epoch 1: loss=3.9752, train=0.0983, test=0.1432
[BP] Epoch 10: loss=3.0848, train=0.2424, test=0.2492
[BP] Epoch 20: loss=2.8041, train=0.2935, test=0.2883
[BP] Epoch 30: loss=2.6019, train=0.3352, test=0.3078
[BP] Epoch 40: loss=2.4193, train=0.3727, test=0.3158
[BP] Epoch 50: loss=2.2631, train=0.4053, test=0.3160
[BP] Epoch 60: loss=2.1134, train=0.4371, test=0.3223
[BP] Epoch 70: loss=1.9686, train=0.4729, test=0.3207
[BP] Epoch 80: loss=1.8724, train=0.4941, test=0.3197
[BP] Epoch 90: loss=1.8161, train=0.5069, test=0.3197
[BP] Epoch 100: loss=1.7897, train=0.5126, test=0.3192
Final test acc: 0.3192
--- DFA ---
[DFA] Epoch 1: loss=4.1736, train=0.0679, test=0.0775
[DFA] Epoch 10: loss=4.0844, train=0.0798, test=0.0819
[DFA] Epoch 20: loss=4.0627, train=0.0840, test=0.0759
[DFA] Epoch 30: loss=4.0466, train=0.0876, test=0.0840
[DFA] Epoch 40: loss=4.0357, train=0.0896, test=0.0862
[DFA] Epoch 50: loss=4.0347, train=0.0909, test=0.0879
[DFA] Epoch 60: loss=4.0298, train=0.0933, test=0.0879
[DFA] Epoch 70: loss=4.0244, train=0.0958, test=0.0883
[DFA] Epoch 80: loss=4.0232, train=0.0939, test=0.0871
[DFA] Epoch 90: loss=4.0219, train=0.0962, test=0.0870
[DFA] Epoch 100: loss=4.0244, train=0.0949, test=0.0875
Final test acc: 0.0875
--- FA ---
[FA] Epoch 1: loss=4.1842, train=0.0639, test=0.0598
[FA] Epoch 10: loss=3.9551, train=0.0978, test=0.0949
[FA] Epoch 20: loss=3.8745, train=0.1103, test=0.1101
[FA] Epoch 30: loss=3.8457, train=0.1160, test=0.1212
[FA] Epoch 40: loss=3.7975, train=0.1235, test=0.1247
[FA] Epoch 50: loss=3.7623, train=0.1290, test=0.1332
[FA] Epoch 60: loss=3.7338, train=0.1341, test=0.1397
[FA] Epoch 70: loss=3.7109, train=0.1404, test=0.1400
[FA] Epoch 80: loss=3.6910, train=0.1426, test=0.1457
[FA] Epoch 90: loss=3.6844, train=0.1436, test=0.1455
[FA] Epoch 100: loss=3.6859, train=0.1440, test=0.1464
Final test acc: 0.1464
============================================================
Seed 123
============================================================
--- BP ---
[BP] Epoch 1: loss=3.9679, train=0.0986, test=0.1439
[BP] Epoch 10: loss=3.0754, train=0.2440, test=0.2501
[BP] Epoch 20: loss=2.8025, train=0.2938, test=0.2812
[BP] Epoch 30: loss=2.5874, train=0.3376, test=0.3021
[BP] Epoch 40: loss=2.4113, train=0.3699, test=0.3104
[BP] Epoch 50: loss=2.2468, train=0.4084, test=0.3160
[BP] Epoch 60: loss=2.1034, train=0.4373, test=0.3209
[BP] Epoch 70: loss=1.9664, train=0.4711, test=0.3212
[BP] Epoch 80: loss=1.8659, train=0.4913, test=0.3208
[BP] Epoch 90: loss=1.8143, train=0.5098, test=0.3201
[BP] Epoch 100: loss=1.7758, train=0.5158, test=0.3218
Final test acc: 0.3218
--- DFA ---
[DFA] Epoch 1: loss=4.1790, train=0.0644, test=0.0808
[DFA] Epoch 10: loss=4.1013, train=0.0738, test=0.0764
[DFA] Epoch 20: loss=4.0720, train=0.0808, test=0.0803
[DFA] Epoch 30: loss=4.0493, train=0.0865, test=0.0845
[DFA] Epoch 40: loss=4.0403, train=0.0866, test=0.0855
[DFA] Epoch 50: loss=4.0321, train=0.0897, test=0.0852
[DFA] Epoch 60: loss=4.0243, train=0.0921, test=0.0856
[DFA] Epoch 70: loss=4.0213, train=0.0924, test=0.0868
[DFA] Epoch 80: loss=4.0207, train=0.0933, test=0.0867
[DFA] Epoch 90: loss=4.0178, train=0.0948, test=0.0875
[DFA] Epoch 100: loss=4.0181, train=0.0932, test=0.0872
Final test acc: 0.0872
--- FA ---
[FA] Epoch 1: loss=4.1971, train=0.0632, test=0.0708
[FA] Epoch 10: loss=4.0477, train=0.0854, test=0.0847
[FA] Epoch 20: loss=3.9867, train=0.0968, test=0.0997
[FA] Epoch 30: loss=3.9504, train=0.1036, test=0.1037
[FA] Epoch 40: loss=3.9204, train=0.1070, test=0.1068
[FA] Epoch 50: loss=3.8915, train=0.1107, test=0.1091
[FA] Epoch 60: loss=3.8680, train=0.1147, test=0.1135
[FA] Epoch 70: loss=3.8517, train=0.1166, test=0.1156
[FA] Epoch 80: loss=3.8433, train=0.1188, test=0.1182
[FA] Epoch 90: loss=3.8342, train=0.1202, test=0.1215
[FA] Epoch 100: loss=3.8330, train=0.1228, test=0.1208
Final test acc: 0.1208
============================================================
Seed 456
============================================================
--- BP ---
[BP] Epoch 1: loss=3.9722, train=0.0978, test=0.1436
[BP] Epoch 10: loss=3.0679, train=0.2433, test=0.2496
[BP] Epoch 20: loss=2.7902, train=0.2983, test=0.2857
[BP] Epoch 30: loss=2.5920, train=0.3374, test=0.3018
[BP] Epoch 40: loss=2.4046, train=0.3747, test=0.3166
[BP] Epoch 50: loss=2.2421, train=0.4090, test=0.3165
[BP] Epoch 60: loss=2.0908, train=0.4420, test=0.3204
[BP] Epoch 70: loss=1.9548, train=0.4750, test=0.3202
[BP] Epoch 80: loss=1.8580, train=0.4973, test=0.3177
[BP] Epoch 90: loss=1.8029, train=0.5128, test=0.3217
[BP] Epoch 100: loss=1.7769, train=0.5179, test=0.3219
Final test acc: 0.3219
--- DFA ---
[DFA] Epoch 1: loss=4.1619, train=0.0684, test=0.0832
[DFA] Epoch 10: loss=4.0780, train=0.0790, test=0.0777
[DFA] Epoch 20: loss=4.0602, train=0.0848, test=0.0813
[DFA] Epoch 30: loss=4.0430, train=0.0885, test=0.0878
[DFA] Epoch 40: loss=4.0391, train=0.0893, test=0.0872
[DFA] Epoch 50: loss=4.0372, train=0.0914, test=0.0834
[DFA] Epoch 60: loss=4.0358, train=0.0919, test=0.0884
[DFA] Epoch 70: loss=4.0340, train=0.0928, test=0.0906
[DFA] Epoch 80: loss=4.0334, train=0.0926, test=0.0879
[DFA] Epoch 90: loss=4.0325, train=0.0935, test=0.0898
[DFA] Epoch 100: loss=4.0329, train=0.0929, test=0.0894
Final test acc: 0.0894
--- FA ---
[FA] Epoch 1: loss=4.2178, train=0.0611, test=0.0534
[FA] Epoch 10: loss=3.9339, train=0.1008, test=0.0999
[FA] Epoch 20: loss=3.8903, train=0.1079, test=0.1125
[FA] Epoch 30: loss=3.8439, train=0.1169, test=0.1138
[FA] Epoch 40: loss=3.8094, train=0.1220, test=0.1228
[FA] Epoch 50: loss=3.7933, train=0.1252, test=0.1240
[FA] Epoch 60: loss=3.7808, train=0.1273, test=0.1275
[FA] Epoch 70: loss=3.7675, train=0.1281, test=0.1252
[FA] Epoch 80: loss=3.7592, train=0.1312, test=0.1307
[FA] Epoch 90: loss=3.7554, train=0.1333, test=0.1311
[FA] Epoch 100: loss=3.7508, train=0.1319, test=0.1310
Final test acc: 0.1310
All results saved to results/cifar100_protocol_validation/results_cifar100.json
--- Frozen baseline on CIFAR-100 ---
Frozen baseline seed=42 (Wed Apr 29 10:19:56 PM CDT 2026)
Device: cuda:0, seed=42, epochs=100, dataset=cifar100
=== BP shallow (ResMLP num_blocks=0), seed=42 ===
n_params: 812900 (812900 trainable)
[BP-shallow] ep 1: test_acc=0.1119
[BP-shallow] ep 10: test_acc=0.1320
[BP-shallow] ep 20: test_acc=0.1286
[BP-shallow] ep 30: test_acc=0.1324
[BP-shallow] ep 40: test_acc=0.1407
[BP-shallow] ep 50: test_acc=0.1555
[BP-shallow] ep 60: test_acc=0.1599
[BP-shallow] ep 70: test_acc=0.1706
[BP-shallow] ep 80: test_acc=0.1738
[BP-shallow] ep 90: test_acc=0.1780
[BP-shallow] ep 100: test_acc=0.1787
FINAL BP-shallow: 0.1787
=== BP frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=42 ===
n_params: 1341284 (812900 trainable)
[BP-frozen] ep 1: test_acc=0.1109
[BP-frozen] ep 10: test_acc=0.1313
[BP-frozen] ep 20: test_acc=0.1252
[BP-frozen] ep 30: test_acc=0.1271
[BP-frozen] ep 40: test_acc=0.1338
[BP-frozen] ep 50: test_acc=0.1557
[BP-frozen] ep 60: test_acc=0.1613
[BP-frozen] ep 70: test_acc=0.1713
[BP-frozen] ep 80: test_acc=0.1751
[BP-frozen] ep 90: test_acc=0.1764
[BP-frozen] ep 100: test_acc=0.1770
FINAL BP-frozen-blocks: 0.1770
=== DFA shallow (ResMLP num_blocks=0), seed=42 ===
n_params: 812900 (812900 trainable)
[DFA-shallow] ep 1: test_acc=0.0914
[DFA-shallow] ep 10: test_acc=0.1120
[DFA-shallow] ep 20: test_acc=0.1130
[DFA-shallow] ep 30: test_acc=0.1198
[DFA-shallow] ep 40: test_acc=0.1170
[DFA-shallow] ep 50: test_acc=0.1211
[DFA-shallow] ep 60: test_acc=0.1248
[DFA-shallow] ep 70: test_acc=0.1203
[DFA-shallow] ep 80: test_acc=0.1248
[DFA-shallow] ep 90: test_acc=0.1254
[DFA-shallow] ep 100: test_acc=0.1255
FINAL DFA-shallow: 0.1255
=== DFA frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=42 ===
n_params: 1341284 (812900 trainable)
[DFA-frozen] ep 1: test_acc=0.0920
[DFA-frozen] ep 10: test_acc=0.1004
[DFA-frozen] ep 20: test_acc=0.1171
[DFA-frozen] ep 30: test_acc=0.1141
[DFA-frozen] ep 40: test_acc=0.1207
[DFA-frozen] ep 50: test_acc=0.1208
[DFA-frozen] ep 60: test_acc=0.1204
[DFA-frozen] ep 70: test_acc=0.1235
[DFA-frozen] ep 80: test_acc=0.1243
[DFA-frozen] ep 90: test_acc=0.1262
[DFA-frozen] ep 100: test_acc=0.1256
FINAL DFA-frozen-blocks: 0.1256
=== ResMLP frozen/shallow baseline summary, seed=42 ===
BP-shallow: 0.1787
BP-frozen: 0.1770
DFA-shallow: 0.1255
DFA-frozen: 0.1256
Compare to trainable 4-block ResMLP (3-seed): BP=0.6147 100ep / 0.585 30ep, DFA=0.306 100ep / 0.301 30ep
Interpretation:
If DFA-frozen ≈ DFA-trainable: blocks are passengers, walk-back parallels ViT
If DFA-frozen << DFA-trainable: ResMLP DFA actually trains the blocks (interesting contrast with ViT)
Frozen baseline seed=123 (Wed Apr 29 10:59:17 PM CDT 2026)
Device: cuda:0, seed=123, epochs=100, dataset=cifar100
=== BP shallow (ResMLP num_blocks=0), seed=123 ===
n_params: 812900 (812900 trainable)
[BP-shallow] ep 1: test_acc=0.1098
[BP-shallow] ep 10: test_acc=0.1309
[BP-shallow] ep 20: test_acc=0.1203
[BP-shallow] ep 30: test_acc=0.1262
[BP-shallow] ep 40: test_acc=0.1415
[BP-shallow] ep 50: test_acc=0.1532
[BP-shallow] ep 60: test_acc=0.1622
[BP-shallow] ep 70: test_acc=0.1725
[BP-shallow] ep 80: test_acc=0.1751
[BP-shallow] ep 90: test_acc=0.1745
[BP-shallow] ep 100: test_acc=0.1756
FINAL BP-shallow: 0.1756
=== BP frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=123 ===
n_params: 1341284 (812900 trainable)
[BP-frozen] ep 1: test_acc=0.1100
[BP-frozen] ep 10: test_acc=0.1328
[BP-frozen] ep 20: test_acc=0.1256
[BP-frozen] ep 30: test_acc=0.1333
[BP-frozen] ep 40: test_acc=0.1411
[BP-frozen] ep 50: test_acc=0.1596
[BP-frozen] ep 60: test_acc=0.1638
[BP-frozen] ep 70: test_acc=0.1720
[BP-frozen] ep 80: test_acc=0.1737
[BP-frozen] ep 90: test_acc=0.1769
[BP-frozen] ep 100: test_acc=0.1777
FINAL BP-frozen-blocks: 0.1777
=== DFA shallow (ResMLP num_blocks=0), seed=123 ===
n_params: 812900 (812900 trainable)
[DFA-shallow] ep 1: test_acc=0.0928
[DFA-shallow] ep 10: test_acc=0.1025
[DFA-shallow] ep 20: test_acc=0.1146
[DFA-shallow] ep 30: test_acc=0.1180
[DFA-shallow] ep 40: test_acc=0.1239
[DFA-shallow] ep 50: test_acc=0.1283
[DFA-shallow] ep 60: test_acc=0.1204
[DFA-shallow] ep 70: test_acc=0.1237
[DFA-shallow] ep 80: test_acc=0.1261
[DFA-shallow] ep 90: test_acc=0.1261
[DFA-shallow] ep 100: test_acc=0.1269
FINAL DFA-shallow: 0.1269
=== DFA frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=123 ===
n_params: 1341284 (812900 trainable)
[DFA-frozen] ep 1: test_acc=0.0916
[DFA-frozen] ep 10: test_acc=0.1060
[DFA-frozen] ep 20: test_acc=0.1167
[DFA-frozen] ep 30: test_acc=0.1125
[DFA-frozen] ep 40: test_acc=0.1153
[DFA-frozen] ep 50: test_acc=0.1237
[DFA-frozen] ep 60: test_acc=0.1217
[DFA-frozen] ep 70: test_acc=0.1254
[DFA-frozen] ep 80: test_acc=0.1239
[DFA-frozen] ep 90: test_acc=0.1254
[DFA-frozen] ep 100: test_acc=0.1257
FINAL DFA-frozen-blocks: 0.1257
=== ResMLP frozen/shallow baseline summary, seed=123 ===
BP-shallow: 0.1756
BP-frozen: 0.1777
DFA-shallow: 0.1269
DFA-frozen: 0.1257
Compare to trainable 4-block ResMLP (3-seed): BP=0.6147 100ep / 0.585 30ep, DFA=0.306 100ep / 0.301 30ep
Interpretation:
If DFA-frozen ≈ DFA-trainable: blocks are passengers, walk-back parallels ViT
If DFA-frozen << DFA-trainable: ResMLP DFA actually trains the blocks (interesting contrast with ViT)
Frozen baseline seed=456 (Wed Apr 29 11:38:34 PM CDT 2026)
Device: cuda:0, seed=456, epochs=100, dataset=cifar100
=== BP shallow (ResMLP num_blocks=0), seed=456 ===
n_params: 812900 (812900 trainable)
[BP-shallow] ep 1: test_acc=0.1073
[BP-shallow] ep 10: test_acc=0.1327
[BP-shallow] ep 20: test_acc=0.1250
[BP-shallow] ep 30: test_acc=0.1303
[BP-shallow] ep 40: test_acc=0.1411
[BP-shallow] ep 50: test_acc=0.1529
[BP-shallow] ep 60: test_acc=0.1651
[BP-shallow] ep 70: test_acc=0.1724
[BP-shallow] ep 80: test_acc=0.1743
[BP-shallow] ep 90: test_acc=0.1757
[BP-shallow] ep 100: test_acc=0.1776
FINAL BP-shallow: 0.1776
=== BP frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=456 ===
n_params: 1341284 (812900 trainable)
[BP-frozen] ep 1: test_acc=0.1073
[BP-frozen] ep 10: test_acc=0.1326
[BP-frozen] ep 20: test_acc=0.1226
[BP-frozen] ep 30: test_acc=0.1276
[BP-frozen] ep 40: test_acc=0.1495
[BP-frozen] ep 50: test_acc=0.1535
[BP-frozen] ep 60: test_acc=0.1645
[BP-frozen] ep 70: test_acc=0.1685
[BP-frozen] ep 80: test_acc=0.1773
[BP-frozen] ep 90: test_acc=0.1777
[BP-frozen] ep 100: test_acc=0.1794
FINAL BP-frozen-blocks: 0.1794
=== DFA shallow (ResMLP num_blocks=0), seed=456 ===
n_params: 812900 (812900 trainable)
[DFA-shallow] ep 1: test_acc=0.0913
[DFA-shallow] ep 10: test_acc=0.1150
[DFA-shallow] ep 20: test_acc=0.1153
[DFA-shallow] ep 30: test_acc=0.1161
[DFA-shallow] ep 40: test_acc=0.1140
[DFA-shallow] ep 50: test_acc=0.1202
[DFA-shallow] ep 60: test_acc=0.1229
[DFA-shallow] ep 70: test_acc=0.1243
[DFA-shallow] ep 80: test_acc=0.1240
[DFA-shallow] ep 90: test_acc=0.1237
[DFA-shallow] ep 100: test_acc=0.1235
FINAL DFA-shallow: 0.1235
=== DFA frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=456 ===
n_params: 1341284 (812900 trainable)
[DFA-frozen] ep 1: test_acc=0.0862
[DFA-frozen] ep 10: test_acc=0.1059
[DFA-frozen] ep 20: test_acc=0.1130
[DFA-frozen] ep 30: test_acc=0.1200
[DFA-frozen] ep 40: test_acc=0.1227
[DFA-frozen] ep 50: test_acc=0.1177
[DFA-frozen] ep 60: test_acc=0.1229
[DFA-frozen] ep 70: test_acc=0.1232
[DFA-frozen] ep 80: test_acc=0.1240
[DFA-frozen] ep 90: test_acc=0.1237
[DFA-frozen] ep 100: test_acc=0.1236
FINAL DFA-frozen-blocks: 0.1236
=== ResMLP frozen/shallow baseline summary, seed=456 ===
BP-shallow: 0.1776
BP-frozen: 0.1794
DFA-shallow: 0.1235
DFA-frozen: 0.1236
Compare to trainable 4-block ResMLP (3-seed): BP=0.6147 100ep / 0.585 30ep, DFA=0.306 100ep / 0.301 30ep
Interpretation:
If DFA-frozen ≈ DFA-trainable: blocks are passengers, walk-back parallels ViT
If DFA-frozen << DFA-trainable: ResMLP DFA actually trains the blocks (interesting contrast with ViT)
=== CIFAR-100 VALIDATION DONE (Thu Apr 30 12:17:41 AM CDT 2026) ===
|
