The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks
Paper • 1803.03635 • Published
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: CastError
Message: Couldn't cast
paper_id: string
overall_verdict: string
evaluated_at: string
note: string
claims: list<item: struct<claim_id: string, description: string, section: string, metric: string, claimed_va (... 129 chars omitted)
child 0, item: struct<claim_id: string, description: string, section: string, metric: string, claimed_value: double (... 117 chars omitted)
child 0, claim_id: string
child 1, description: string
child 2, section: string
child 3, metric: string
child 4, claimed_value: double
child 5, observed_value: double
child 6, tolerance: double
child 7, tolerance_kind: string
child 8, delta: double
child 9, allowed: double
child 10, verdict: string
to
{'id': Value('string'), 'description': Value('string'), 'section': Value('string'), 'quote': Value('string'), 'metric': Value('string'), 'claimed_value': Value('float64'), 'tolerance': Value('float64'), 'tolerance_kind': Value('string'), 'tolerance_rationale': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1837, in _prepare_split_single
writer.write_table(table)
~~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 764, in write_table
self.write_rows_on_file() # in case there are buffered rows to write first
~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 663, in write_rows_on_file
self._write_table(table)
~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
pa_table = table_cast(pa_table, self._schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
paper_id: string
overall_verdict: string
evaluated_at: string
note: string
claims: list<item: struct<claim_id: string, description: string, section: string, metric: string, claimed_va (... 129 chars omitted)
child 0, item: struct<claim_id: string, description: string, section: string, metric: string, claimed_value: double (... 117 chars omitted)
child 0, claim_id: string
child 1, description: string
child 2, section: string
child 3, metric: string
child 4, claimed_value: double
child 5, observed_value: double
child 6, tolerance: double
child 7, tolerance_kind: string
child 8, delta: double
child 9, allowed: double
child 10, verdict: string
to
{'id': Value('string'), 'description': Value('string'), 'section': Value('string'), 'quote': Value('string'), 'metric': Value('string'), 'claimed_value': Value('float64'), 'tolerance': Value('float64'), 'tolerance_kind': Value('string'), 'tolerance_rationale': Value('string')}
because column names don't match
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1839, in _prepare_split_single
raise DatasetGenerationCastError.from_cast_error(
...<4 lines>...
)
datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 5 new columns ({'claims', 'note', 'overall_verdict', 'evaluated_at', 'paper_id'}) and 9 missing columns ({'tolerance', 'quote', 'id', 'tolerance_rationale', 'description', 'metric', 'claimed_value', 'tolerance_kind', 'section'}).
This happened while the json dataset builder was generating data using
hf://datasets/Kowshik71/repro-lottery-ticket/lenet/claims.json (at revision a64e737c68c90d9dfd0d1752025f74079d026f78), ['hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/cifar/claims.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/claim_result_matrix.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/claims.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/experiment.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/experiment_manifest.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/logs/run.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/provenance.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed0.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed1.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed2.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed3.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed4.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed0.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed1.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed2.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed3.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed4.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/summary.json'], ['hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/cifar/claims.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/claim_result_matrix.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/claims.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/experiment.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/experiment_manifest.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/logs/run.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/provenance.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed0.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed1.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed2.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed3.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/claim_result_matrix-seed4.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed0.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed1.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed2.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed3.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/seeds/metrics-seed4.json', 'hf://datasets/Kowshik71/repro-lottery-ticket@a64e737c68c90d9dfd0d1752025f74079d026f78/lenet/summary.json']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1858, in _prepare_split_single
num_examples, num_bytes = writer.finalize()
~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 781, in finalize
self.write_rows_on_file()
~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 663, in write_rows_on_file
self._write_table(table)
~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
pa_table = table_cast(pa_table, self._schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
paper_id: string
overall_verdict: string
evaluated_at: string
note: string
claims: list<item: struct<claim_id: string, description: string, section: string, metric: string, claimed_va (... 129 chars omitted)
child 0, item: struct<claim_id: string, description: string, section: string, metric: string, claimed_value: double (... 117 chars omitted)
child 0, claim_id: string
child 1, description: string
child 2, section: string
child 3, metric: string
child 4, claimed_value: double
child 5, observed_value: double
child 6, tolerance: double
child 7, tolerance_kind: string
child 8, delta: double
child 9, allowed: double
child 10, verdict: string
to
{'id': Value('string'), 'description': Value('string'), 'section': Value('string'), 'quote': Value('string'), 'metric': Value('string'), 'claimed_value': Value('float64'), 'tolerance': Value('float64'), 'tolerance_kind': Value('string'), 'tolerance_rationale': Value('string')}
because column names don't match
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1869, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
id string | description string | section string | quote string | metric string | claimed_value float64 | tolerance float64 | tolerance_kind string | tolerance_rationale string |
|---|---|---|---|---|---|---|---|---|
S1 | Structural: Conv-2 as described in Appendix H.1 and Figure 2 has 4.3M weights (biases excluded). Verifies the reproduction built the architecture the paper describes before any claim about it is scored. | Figure 2, 'All/Conv Weights' row, Conv-2 column | All/Conv Weights 266K | 4.3M / 38K | 2.4M / 260K | 1.7M / 1.1M | 274K / 270K | 20.0M | conv2_total_weights | 4,300,000 | 50,000 | abs | Figure 2 states the count to two significant figures, so +/-50,000 is the half-width of the rounding band on '4.3M'. Deterministic arithmetic, not a measurement: computed pre-run as 3*3*3*64 + 3*3*64*64 + 16384*256 + 256*256 + 256*10 = 4,300,992. Failing this means padding, pooling, or the flatten size is wrong and eve... |
S2 | Structural: Conv-4 has 2.4M weights (biases excluded). Second, independent confirmation of the same 'same'-padding + pool-after-every-module geometry used for Conv-2. | Figure 2, 'All/Conv Weights' row, Conv-4 column | All/Conv Weights 266K | 4.3M / 38K | 2.4M / 260K | 1.7M / 1.1M | 274K / 270K | 20.0M | conv4_total_weights | 2,400,000 | 50,000 | abs | Same two-significant-figure rounding band as S1. Computed pre-run as 259,776 conv + 8192*256 + 256*256 + 256*10 = 2,425,024. This is the claim that discriminates the geometry: 'valid' padding gives 1,147,072 and dropping the final pool gives 8,716,480, both far outside the band. |
S3 | Structural: Conv-6's convolutional weight count is 1.1M. Paired with S4 to localise the paper's Conv-6 inconsistency to the fully-connected/total accounting rather than to the convolutional stack. | Figure 2, 'All/Conv Weights' row, Conv-6 column | All/Conv Weights 266K | 4.3M / 38K | 2.4M / 260K | 1.7M / 1.1M | 274K / 270K | 20.0M | conv6_conv_weights | 1,100,000 | 50,000 | abs | Same rounding band as S1/S2. Computed pre-run as 1,728+36,864+73,728+147,456+294,912+589,824 = 1,144,512, which is inside the band. Expected to PASS, and its passing is what makes the S4 failure diagnostic rather than a sign of a broken implementation. |
S4 | Structural: Conv-6's stated total weight count is 1.7M. REGISTERED AS AN EXPECTED FAILURE. The architecture Appendix H.1 and Figure 2 describe cannot have 1.7M weights; the reproduction will build 2,261,184 and report the discrepancy as a defect in the paper. | Figure 2, 'All/Conv Weights' row, Conv-6 column; propagated into Section 3 and Appendix H.6 | All/Conv Weights ... 1.7M / 1.1M ... and Appendix H.6: 'fully-connected layers comprise 99%, 89%, and 35% of the parameters in Conv-2, Conv-4, and Conv-6' | conv6_total_weights | 1,700,000 | 50,000 | abs | Held to exactly the same two-significant-figure band as S1-S3; no special pleading. Pre-run arithmetic already shows this FAILS: the described architecture computes to 2,261,184, missing by 561,184 (11x the tolerance). Reaching 1.7M would require an fc1 input of 1,903.9 units, which is not an integer and corresponds to... |
S5 | Structural: the pruning schedule (conv 10%, fc 20%, output 10% per round, layer-wise, applied to surviving weights) places Conv-2's 11th pruning round at Pm=8.8%, the sparsity the paper attaches to its Conv-2 speedup claim. | Figure 2 'Pruning Rate' row + Appendix H.1 (output layer at half the fc rate) + Section 3 Pm labels | Pruning Rate fc20% | conv10% fc20% | conv10% fc20% | conv15% fc20% ... and 'The output layer is pruned at half of the rate of the fully-connected layers for the reasons described in Appendix G.' | conv2_pm_percent_round11 | 8.8 | 0.1 | abs | Deterministic arithmetic; computed pre-run as 8.808%. The 0.1pp band absorbs integer per-layer survivor counts (floor/round shifts Pm by <0.01pp at these layer sizes) while still catching the realistic implementation bugs: using conv15% instead of conv10% moves this rung by ~1.5pp, and pruning the output layer at the f... |
S6 | Structural: the same schedule places Conv-4's 12th pruning round at Pm=9.2%, the sparsity the paper attaches to its Conv-4 speedup claim. Conv-4's other quoted rung, Pm=11.1%, is round 11. | Figure 2 'Pruning Rate' row + Appendix H.4 + Section 3 Pm labels | we select an iterative convolutional pruning rate of 10% for Conv-2, 10% for Conv-4, and 15% for Conv-6. | conv4_pm_percent_round12 | 9.2 | 0.1 | abs | Same band and same justification as S5; computed pre-run as 9.184%. Independently corroborated three ways from the paper's own text: round 11 gives 11.055% (the paper's Pm=11.1% accuracy rung) and round 15 gives 5.365% (the Pm=5.4% rung quoted in Appendix H.2). Under conv15% neither lands, which is the arithmetic that ... |
S7 | Structural: the same schedule at Conv-6's conv15%/fc20%/out10% rates places its 10th pruning round at Pm=15.1%, the sparsity the paper attaches to its Conv-6 speedup claim. REGISTERED AS AN EXPECTED FAILURE, at the same tolerance as S5 and S6. | Figure 2 'Pruning Rate' row + Appendix H.4 + Section 3 Pm labels | 2.5x for Conv-6 (P_m=15.1%) ... we select an iterative convolutional pruning rate of ... 15% for Conv-6. | conv6_pm_percent_round10 | 15.1 | 0.1 | abs | Deliberately held to the same 0.1pp band as S5/S6 rather than being widened to fit. Pre-run arithmetic gives 15.295% on the architecturally-forced 2,261,184-weight network, missing by 0.195pp. The offset is systematic, not random: Conv-6's other quoted rung, Pm=26.4%, also lands 0.21pp high (26.613% at round 7), and bo... |
C1 | Winning tickets learn faster on Conv-2: the best winning ticket reaches minimum validation loss 3.5x faster than the unpruned network. Measured as max over swept rounds of [early-stopping iteration of round 0] / [early-stopping iteration of round k], on seed-averaged curves. The paper attaches this maximum to Pm=8.8% (... | Section 3, 'Finding winning tickets' paragraph (Figure 5, top, solid lines, no dropout) | Winning tickets reach minimum validation loss at best 3.5x faster for Conv-2 (P_m=8.8%), 3.5x for Conv-4 (P_m=9.2%), and 2.5x for Conv-6 (P_m=15.1%). | conv2_max_early_stop_speedup | 3.5 | 1.3 | abs | The paper reports this exact quantity twice, at the same Pm=8.8% rung, and disagrees with itself: 3.5x in Section 3 (5 trials, Figure 5) and 4.8x in Appendix H.2 (3 trials, learning-rate sweep, same selected lr 2e-4). The paper's own test-retest spread on this metric is therefore 1.3x, and no tolerance narrower than th... |
C2 | Winning tickets generalise better on Conv-2: test accuracy at the early-stopping iteration improves by at best 3.4 percentage points over the unpruned network. Measured as max over swept rounds of [mean test accuracy at early stop, round k] - [mean test accuracy at early stop, round 0]. The paper attaches this maximum ... | Section 3, 'Finding winning tickets' paragraph (Figure 5, top-middle, solid lines, no dropout) | Test accuracy improves at best 3.4 percentage points for Conv-2 (P_m=4.6%), 3.5 for Conv-4 (P_m=11.1%), and 3.3 for Conv-6 (P_m=26.4%). | conv2_max_test_acc_delta_pp | 3.4 | 1.5 | abs | Budget: the paper's own internal spread on this quantity is up to 0.9pp (Section 3 reports 3.4pp test for Conv-2 while Appendix H.2 reports 3.3pp validation, and for Conv-6 the gap between the two is 3.3pp vs 2.4pp); binomial noise on a 10,000-example test set at ~70% accuracy contributes ~0.65pp to a difference of two... |
C3 | Winning tickets learn faster on Conv-4: the best winning ticket reaches minimum validation loss 3.5x faster than the unpruned network. Same estimator as C1. The paper attaches this maximum to Pm=9.2% (round 12). | Section 3, 'Finding winning tickets' paragraph (Figure 5, top, solid lines, no dropout) | Winning tickets reach minimum validation loss at best 3.5x faster for Conv-2 (P_m=8.8%), 3.5x for Conv-4 (P_m=9.2%), and 2.5x for Conv-6 (P_m=15.1%). | conv4_max_early_stop_speedup | 3.5 | 1.3 | abs | Same band and same reasoning as C1, applied to the same metric on a different architecture. Corroborated independently for Conv-4: Appendix H.2 reports 4.27x at Pm=11.1% under the selected lr 3e-4, a 0.77x disagreement with the main body's 3.5x. +/-1.3 accepts 2.2x-4.8x, which covers the paper's own 4.27x and fails a n... |
C4 | Winning tickets generalise better on Conv-4: test accuracy at the early-stopping iteration improves by at best 3.5 percentage points. Same estimator as C2. The paper attaches this maximum to Pm=11.1% (round 11). | Section 3, 'Finding winning tickets' paragraph (Figure 5, top-middle, solid lines, no dropout) | Test accuracy improves at best 3.4 percentage points for Conv-2 (P_m=4.6%), 3.5 for Conv-4 (P_m=11.1%), and 3.3 for Conv-6 (P_m=26.4%). | conv4_max_test_acc_delta_pp | 3.5 | 1.5 | abs | Same band and same budget as C2. Corroborated for Conv-4 by Appendix H.2, which reports a 3.7pp validation improvement at Pm=5.4% under the selected lr, 0.2pp from the main body's 3.5pp. Accepts 2.0-5.0pp; a null result fails. |
C5 | Winning tickets learn faster on Conv-6: the best winning ticket reaches minimum validation loss 2.5x faster than the unpruned network. Same estimator as C1. The paper attaches this maximum to Pm=15.1%, which is round 10 on the reconstructed ladder (see S7 - the rung is addressed by round index, not by the printed Pm). | Section 3, 'Finding winning tickets' paragraph (Figure 5, top, solid lines, no dropout) | Winning tickets reach minimum validation loss at best 3.5x faster for Conv-2 (P_m=8.8%), 3.5x for Conv-4 (P_m=9.2%), and 2.5x for Conv-6 (P_m=15.1%). | conv6_max_early_stop_speedup | 2.5 | 0.9 | abs | The same 37% relative band used for C1 and C3 (1.3/3.5), rescaled to the smaller claimed value: 2.5 * 0.37 = 0.93, rounded to 0.9. This also matches the +/-0.9 band the already-published Lenet half of this reproduction used for the structurally identical 2.51x speedup claim, so the two halves are held to one standard. ... |
C6 | Winning tickets generalise better on Conv-6: test accuracy at the early-stopping iteration improves by at best 3.3 percentage points. Same estimator as C2. The paper attaches this maximum to Pm=26.4%, which is round 7 on the reconstructed ladder (addressed by round index, see S7). | Section 3, 'Finding winning tickets' paragraph (Figure 5, top-middle, solid lines, no dropout) | Test accuracy improves at best 3.4 percentage points for Conv-2 (P_m=4.6%), 3.5 for Conv-4 (P_m=11.1%), and 3.3 for Conv-6 (P_m=26.4%). | conv6_max_test_acc_delta_pp | 3.3 | 1.5 | abs | Same band and budget as C2 and C4. Conv-6 is where the paper disagrees with itself most on this metric: Section 3 reports 3.3pp (test, 5 trials) while Appendix H.2 reports 2.4pp (validation, 3 trials) - a 0.9pp internal spread, which is the empirical basis for the 1.5pp band across C2/C4/C6. Accepts 1.8-4.8pp; a null r... |
C7 | The sparsity floor: all three convolutional networks stay above their unpruned average test accuracy for every rung with Pm>2%. Operationalised as the Pm of the FIRST rung at which seed-averaged test accuracy at early stop drops below the unpruned average (first-crossing semantics, so a noisy non-monotone curve cannot ... | Section 3, 'Finding winning tickets' paragraph | All three networks remain above their original average test accuracy when P_m>2%. | worst_arch_accuracy_crossover_pm_percent | 2 | 1.5 | abs | The crossover is quantised to the pruning ladder, which near 2% has ~0.4-0.5pp spacing (Conv-2: 2.965, 2.390, 1.928, 1.557), so +/-1.5 is about three rungs either side - the minimum that is not dominated by quantisation, given that accuracy curves are flat and noisy where they cross the baseline and the crossing rung i... |
C8 | The accuracy gain is generalisation, not optimisation: at the final training iteration (20,000 / 25,000 / 30,000 for Conv-2/4/6) training accuracy reaches ~100% for every rung with Pm>=2%, so winning tickets' higher test accuracy cannot be explained by better fitting of the training set. Operationalised as the minimum,... | Section 3; backing graphs in Appendix D, Figure 13; restated in the Figure 5 caption | at iteration 20,000 for Conv-2, 25,000 for Conv-4, and 30,000 for Conv-6 (the iterations corresponding to the final training iteration for the original network), training accuracy reaches 100% for all networks when P_m>=2% (Appendix D, Figure 13) and winning tickets still maintain higher test accuracy | min_final_train_acc_pct_above_2pct | 100 | 2 | abs | The paper itself hedges the figure: the Figure 5 caption writes 'training accuracy ~100% for P_m>=2%' and Appendix D says '100% in all cases for all but the most heavily pruned networks'. A tolerance tighter than the paper's own '~' would be scoring a precision the paper does not assert. The statistic is a minimum over... |
C9 | The initialisation matters, not just the structure: at the sparsest swept rung, winning tickets retain higher test accuracy at early stop than the same masks with randomly resampled initialisations. Registered as a SIGN TEST - the paper makes no numeric reinitialisation claim for Conv-2/4/6, so no magnitude is asserted... | Section 3, 'Random reinitialization' paragraph (Figure 5, dashed lines) | These networks again take increasingly longer to learn upon continued pruning. Just as with Lenet on MNIST (Section 2), test accuracy drops off more quickly for the random reinitialization experiments. | num_archs_ticket_beats_reinit_at_sparsest_rung | 3 | 0.5 | abs | Section 3 states this direction for the convolutional networks but never quantifies it (unlike Section 2, which gives 2.51x and 'half a percentage point' for Lenet), so registering any magnitude would be inventing a number the paper does not contain. A sign test is the faithful encoding: claimed_value 3.0 with +/-0.5 p... |
null | null | null | null | null | null | null | null | null |
C1 | Winning tickets learn faster: at Pm=21.1% the early-stopping iteration (minimum validation loss) occurs 38% earlier than for the unpruned network. | Section 2, 'Iterative pruning' (Figure 4a, left) | The winning tickets learn faster as P_m decreases from 100% to 21%, at which point early-stopping occurs 38% earlier than for the original network. | early_stop_reduction_pct_pm21_1 | 38 | 15 | abs | The paper averages 5 trials; this is 1 seed. Early stopping is the argmin of a validation loss sampled every 100 iterations, which is a noisy statistic. +/-15pp accepts a 23-53% speedup: wide enough for single-seed noise, still failing if pruning does not speed learning up or slows it down. |
C2 | Test accuracy at early stop improves by more than 0.3 percentage points at Pm=13.5% relative to the unpruned network. | Section 2, 'Iterative pruning' (Figure 4a, middle) | Test accuracy increases with pruning, improving by more than 0.3 percentage points when P_m = 13.5%; after this point, accuracy decreases, returning to the level of the original network when P_m = 3.6%. | test_acc_delta_pp_pm13_5 | 0.3 | 0.3 | abs | The paper states a lower bound ('more than 0.3 pp') averaged over 5 trials; the gym scores two-sided, so the stated magnitude is registered as a point value with +/-0.3pp (accepts 0.0-0.6pp). Fails if pruning to 13.5% does not improve accuracy at all. |
C3 | By Pm=3.6% the accuracy gain is gone: test accuracy returns to the level of the original unpruned network. | Section 2, 'Iterative pruning' (Figure 4a, middle) | after this point, accuracy decreases, returning to the level of the original network when P_m = 3.6%. | test_acc_delta_pp_pm3_6 | 0 | 0.3 | abs | 'Returning to the level of the original network' means a delta of zero; +/-0.3pp is the same band used for C2 so the two halves of the same sentence are held to the same standard. |
C4 | The original initialisation matters: at Pm=21% the winning ticket reaches minimum validation loss 2.51x faster than the same mask with a resampled initialisation. | Section 2, 'Random reinitialization' (Figure 4a, orange) | When P_m = 21%, the winning ticket reaches minimum validation loss 2.51x faster than when reinitialized and is half a percentage point more accurate. | speedup_ticket_vs_random_pm21_1 | 2.51 | 0.9 | abs | A ratio of two noisy argmins, here from 1 seed and 1 reinitialisation draw against the paper's 5 trials x 3 reinit draws. +/-0.9 accepts 1.61x-3.41x, which still fails if the ticket is not meaningfully faster than a random reinit (the paper's central control). |
C5 | At Pm=21% the winning ticket is about half a percentage point more accurate than the same mask randomly reinitialised. | Section 2, 'Random reinitialization' (Figure 4a, middle) | When P_m = 21%, the winning ticket reaches minimum validation loss 2.51x faster than when reinitialized and is half a percentage point more accurate. | acc_gap_ticket_minus_random_pp_pm21_1 | 0.5 | 0.4 | abs | Accepts a 0.1-0.9pp advantage for the winning ticket on a single reinit draw; fails if the resampled initialisation matches or beats the ticket. |
C6 | Winning tickets generalise better, not just optimise faster: at iteration 50,000 they still improve test accuracy by up to 0.35 percentage points over the unpruned network. | Section 2, 'Iterative pruning' (Figure 4b) | However, at iteration 50,000 (Figure 4b), iteratively-pruned winning tickets still see a test accuracy improvement of up to 0.35 percentage points in spite of the fact that training accuracy reaches 100% for nearly all networks. | max_test_acc_delta_pp_final_iter | 0.35 | 0.3 | abs | 'Up to' means the maximum over pruning levels, so this statistic is upward-biased on any single run; +/-0.3pp accepts 0.05-0.65pp and fails if no pruning level beats the unpruned network at the final iteration. |
C7 | Structural check of the pruning schedule: pruning 20% of surviving weights per round in each fully-connected layer and 10% in the output layer, layer-wise, leaves Pm=21.1% of the 266,200 weights after 7 rounds. | Section 2 (Pm labels) + Appendix G.1 / Figure 2 ('All Weights 266K') | we perform the lottery ticket experiment iteratively with a pruning rate of 20% per iteration (10% for the output layer) ... Each layer of the network is pruned independently. | pm_percent_pm21_1 | 21.1 | 0.05 | abs | Deterministic arithmetic, not a measurement: it verifies the harness implements the paper's pruning schedule and that biases are excluded (266,200 weights). A tight 0.05pp tolerance is therefore appropriate; a mismatch means the implementation is wrong and the other six claims are untrustworthy. |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
null | null | null | null | null | null | null | null | null |
Replicating The Lottery Ticket Hypothesis
An independent, pre-registered replication of both halves of:
Jonathan Frankle and Michael Carbin. The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks. ICLR 2019. arXiv:1803.03635 · doi:10.48550/arXiv.1803.03635
All credit for the research is the original authors'. This repository is a replication — it contains the reproducer's own harness and evidence, and redistributes no code, data, or weights of the authors'.
This is a replication, not a reproduction, in the ACM sense ("Results Replicated"): the experiments were re-implemented from the paper's text by a different team on different artifacts. The authors' code was not used.
lenet/ — LeNet-300-100 / MNIST |
cifar/ — Conv-2/4/6 / CIFAR-10 |
|
|---|---|---|
| status | complete, 5 seeds | pre-registered, runs in flight |
| claims | 7 | 16 |
| result | 6/7 replicated on the 5-seed mean | pending |
| boundary | rootless podman, contained | Kaggle T4, recorded as external: |
What is actually being claimed here
Every number below was registered — with its verbatim quote from the paper, a tolerance, and a written rationale for that tolerance — before the run that tests it existed. Nothing was retuned after seeing a result.
LeNet/MNIST (done). Six of seven claims replicate on the mean of five seeds.
The one that does not is the paper's "38% earlier early stopping" magnitude: it
came out at 14.7% over three seeds and 21.4% over five, against a pre-registered
tolerance of 38±15. We report it as not replicated at this sample size, and
specifically do not claim the paper is wrong — the quantity is a ratio of two
individually noisy single runs. Details, including a correction to our own n=3
diagnosis, are in lenet/REPRODUCTION_NOTES.md.
CIFAR-10 (running). Sixteen claims, of which two are registered as
predicted failures before any run: the paper's stated Conv-6 parameter count
(1.7M) contradicts the architecture its own appendix describes (2,261,184), and
that propagates into one of its printed sparsity labels. Three independent lines
of evidence are laid out in cifar/PREREGISTRATION.md.
Predicting a failure in public, before running, is the point — it is what makes
the eventual result informative rather than post-hoc.
Why both halves live in one repository
They are one paper. Splitting a replication across repositories by experiment makes the claim ledger harder to read than the paper it is checking.
Beyond the paper's own question
USE_CASES.md lists five things this work is useful for that the
original paper was not written to serve — a measured compression schedule with a
located accuracy cliff, a CI canary for silent training-pipeline regressions, a
way to size the seed count a convergence-speed claim needs, a transferable
template for auditing anyone's numeric claims, and a pre-submission
parameter-count check that catches a class of error peer review reliably misses.
Each points at evidence in this repo, and the caveats are stated with them.
Known defects in this replication
Errata are published, not patched away. See
lenet/ERRATUM.md: two of the seven LeNet claims are
one-sided in the paper's wording ("more than 0.3", "up to 0.35") and were
registered two-sided. Rescored correctly the verdicts are unchanged and the 6/7
headline stands — but the registration was right by luck, not construction, and
under a different draw it would have produced a false failure. The scorer now
supports one-sided and interval claims; the original registration is left
unedited, because a pre-registration rewritten after results exist is worthless.
Honest results
Every registered claim is published with its verdict — replicated, not replicated, or inconclusive. A failure is a real, reportable result and is never dropped. Where our own earlier analysis turned out to be wrong (the n=3 variance diagnosis), the correction is published beside it rather than replacing it.
Reproducing this replication
Each half documents its own environment and commands. Both were produced with paper-repro-gym, which fixes claims and tolerances before a run, scores against them afterwards, and records the containment boundary honestly — including when a run happened on external compute it did not control.
See CITATION.cff to cite both this replication and the
original paper.
- Downloads last month
- 54