PaperlayerView exact revision on arXiv

Source-linked AI summary

LESS: Lightweight Evolutionary Supernet Search in Minutes

Aviral Gandhi, Jinglue Xu, Jialong Li, Hitoshi Iba

arXiv:2610.01468v1cs.NEcs.LG

TL;DR

Low-cost NAS must both encounter strong architectures and reliably select them using shared-weight evaluations. LESS uses brief candidate-conditioned updates to adapt shared weights while scoring proposed architectures. It completes NAS-Bench-201/CIFAR-10 search in 409 seconds and transfers its configuration to other datasets and DARTS, while calibration primarily improves selection.

  • Problem

    Low-cost NAS must both encounter strong architectures and reliably select them using shared-weight evaluations, while low-cost evaluation strategies trade task feedback against training cost.

    Source: §1 · p. 1

  • Method

    LESS combines brief fair supernet training with candidate-conditioned hard-path updates that adapt shared weights while scoring only the proposed architecture.

    Source: §1 · pp. 1–2

  • Results

    LESS completes NAS-Bench-201/CIFAR-10 search in 409 seconds, transfers without search-hyperparameter retuning to other datasets and DARTS, and improves the returned architecture despite little change in the best visited.

    Source: §6 · p. 9

  • Takeaways & Limitations

    A few balanced, task-dependent updates substantially improve the architecture returned by a low-cost search even when the best architecture visited changes little.

    Source: §6 · p. 9

    Transfer experiments do not establish universal ranking reliability; DARTS is evaluated with only three searched architectures per dataset.

    Source: §5 · p. 9

Abstract

from arXiv · show

Low-cost NAS must both explore high-performing architectures and identify them reliably, yet reducing evaluation cost often weakens the fidelity of candidate comparisons. Training-free methods reduce evaluation cost by replacing learned task feedback with proxy signals measured at initialization. We introduce LESS (Lightweight Evolutionary Supernet Search), a data-driven method that combines a brief fair hard-path warm-up with discrete search under a single CMA-ES distribution. Each proposal is evaluated as its decoded hard genotype after six candidate-conditioned supernet updates. On NAS-Bench-201, LESS achieves \(93.189\pm0.467\%\) CIFAR-10 test accuracy in 409.1 seconds, coming within 0.04 percentage points of FairNAS using approximately \(1/24\) of its source-reported search time. Matched controls show that calibration improves selected validation accuracy by \(0.577\) percentage points while changing best-visited accuracy by only \(0.054\) points, indicating that its primary effect is to reduce selection regret. The frozen configuration transfers without tuning to CIFAR-100 and ImageNet16-120 with \(69.615\pm1.139\%\) and \(43.720\pm1.697\%\) accuracy. Applied without tuning to the larger DARTS space, LESS achieves \(96.95\pm0.14\%\) on CIFAR-10 and \(82.43\pm0.80\%\) on CIFAR-100, with each search completing in approximately 43.5 minutes on a single GPU. Together, these results show that short, balanced, data-dependent updates enable competitive neural architecture search across datasets and search spaces within minutes.

1 INTRODUCTION

LESS addresses low-cost NAS as both a search problem and a reliable-selection problem, using brief task-dependent candidate evaluation within an evolutionary search pipeline. Its calibration improves selected validation accuracy more than it changes best-visited accuracy, and the method is applied across datasets and search spaces without retuning core settings.

1 INTRODUCTION

  • Weight-sharing NAS reduces training cost, but must both encounter strong architectures and reliably select them using shared-weight evaluations.Earlier reinforcement-learning and evolutionary systems required hundreds to thousands of accelerator days because candidates were trained from scratch.Source: §1 · p. 1
  • Fair single-path methods preserve learned task feedback but require substantial supernet training, while training-free proxies are fast but their rankings vary across spaces and datasets.Low-fidelity estimators trade limited training for stronger task feedback, but typically require more optimization or separate performance estimation per candidate.Source: §1 · p. 1
  • LESS evaluates proposed architectures with brief, balanced candidate-conditioned updates, retaining only the winning candidate’s updated state for subsequent search.Candidates start from the same weights and use shared data; only the proposed architecture is scored.Source: §1 · pp. 1–2
  • Calibration improves selected validation accuracy by 0.577 percentage points while changing best-visited accuracy by only 0.054 points across 24 paired NAS-Bench-201/CIFAR-10 searches.The authors interpret this pattern as reduced selection regret: the gap between the best architecture encountered and the architecture returned.Source: §1 · p. 2
  • LESS combines evolutionary search with an efficient evaluator and is applied across datasets and search spaces without retuning core search hyperparameters.The contribution framing also separates architecture quality encountered from architecture quality returned.Source: §1 · p. 2, Figure 1 · p. 2

2 RELATED WORK

Related work reduces NAS cost through weight sharing, fair supernet training, training-free proxies, low-fidelity estimation, and evolutionary search. LESS combines candidate-conditioned fair calibration with hard-genotype evaluation and winner-state inheritance.

Weight sharing and fair supernets

  • FairNAS balances hard-path operation exposure, while LESS applies fairness locally during candidate comparison using isolated updates from a common parent state and retaining only the winner’s state.This contrasts with Supernet Shifting, which accumulates updates into one supernet that specializes toward frequently sampled regions.Source: §2 · p. 2

Low-cost evaluation and architecture selection

  • Training-free NAS ranks architectures from signals computed at or near initialization, while low-fidelity estimators use limited training or learned performance predictors for stronger task feedback.Examples include NASWOT activation-pattern diversity and TE-NAS neural-tangent-kernel conditioning with linear-region counts.Source: §2 · p. 2
  • Candidate-conditioned fair evaluation groups complementary hard paths to cover each eligible operation once per edge, averages their gradients, and scores only the candidate after BatchNorm refresh.The refreshed scoring copy is discarded; the saved pre-refresh state is inherited only if the candidate wins the epoch.Source: Figure 2 · p. 3
  • LESS uses bounded task-dependent updates followed by evaluation of the exact hard genotype to address architecture-selection reliability.DARTS-PT shows that optimizing a supernet does not guarantee reliable architecture extraction.Source: §2 · p. 3

Evolutionary search

  • LESS combines CMA-ES exploration with candidate-conditioned fair calibration and persistent winner-state inheritance, rather than using a pretrained one-shot model as a fixed fitness estimator.Evolutionary NAS directly optimizes discrete architectures but is expensive when each proposal is trained independently.Source: §2 · p. 3

3 METHOD

LESS combines discrete CMA-ES search with candidate-conditioned calibration from a shared parent state, then selects the final architecture by normalized Borda aggregation.

3.3 HARD-PATH SCORING, INHERITANCE, AND FINAL SELECTION

  • The epoch winner’s calibrated state becomes the next parent, while cumulative normalized Borda ranks determine the architecture returned.The returned genotype need not be the epoch winner.Source: §3 · p. 3, §3.3 · p. 5

3.1 ARCHITECTURE REPRESENTATION AND DECODING

  • Edge-operation logits parameterize discrete choices; before decoding, probability mass is shifted from none and skip to the remaining operations.The fixed redistribution rates are rnone = 0.55 and rskip = 0.50.Source: §3.1 · p. 3
  • The operation-mass redistribution assumes none and skip retain their standard null and identity semantics; extending the prior to macro-level and transformer spaces remains future work.Source: §3.1 · p. 4
  • A deterministic decoder enforces search-space topology constraints, and duplicate genotypes are removed before forming a population of 12 distinct non-incumbent candidates.A deterministic fallback preserves population size if the bounded proposal budget is exhausted.Source: §3.1 · p. 4

3.2 CANDIDATE-CONDITIONED FAIR CALIBRATION

  • A 25-epoch balanced hard-path warm-up initializes the parent; each candidate then receives six updates from its own copy of that shared state.For NAS-Bench-201, each candidate is included in a fair group that executes every eligible operation once per edge, with complementary paths included in the update but not the score.Source: §3.2 · pp. 4–5

3.3 HARD-PATH SCORING, INHERITANCE, AND FINAL SELECTION

  • Candidates are scored by validation cross-entropy on the decoded hard genotype after a separate BatchNorm refresh; the refreshed scoring copy is discarded.CMA-ES updates from proposal losses, while the best member of the population including the incumbent supplies the continuing search state.Source: §3.3 · p. 5

4 RESULTS

LESS is evaluated on NAS-Bench-201 and the DARTS space, including matched controls, component ablations, dataset transfer, and accuracy–compute scaling. Its results include competitive test accuracy with frozen search settings and show that candidate selection and search configuration affect returned architectures.

4.1 EVALUATION SETUP

  • The frozen NAS-Bench-201 protocol uses 25 warm-up epochs, six calibration updates, 12 proposals plus an incumbent, and 60 search epochs; all search times use one RTX 4090.Training data support warm-up, calibration, and BatchNorm refresh; disjoint validation data score candidates, and held-out search validation determines the returned architecture.Source: §4.1 · p. 5

4.2 NAS-BENCH-201 ACCURACY AND SEARCH COST

  • On NAS-Bench-201, LESS’s test accuracy is within 0.041 percentage points of FairNAS; source-reported search times are 409 and 9,845 seconds, respectively, using different hardware and implementations.These timings do not establish hardware-normalized speedups.Source: §4.2 · p. 6
  • Calibration lowers selection regret from 1.275 to 0.753 points while best-visited validation accuracy changes only from 90.676% to 90.730%.Calibrated trajectories visit fewer distinct genotypes on average (124.9 versus 147.0) but include more candidate occurrences above the global 95th and 99th percentiles.Source: §4.2 · p. 6

4.3 COMPONENT ABLATIONS

  • At W = 25, K = 1 reaches 90.131% mean selected validation accuracy, while K = 6 reaches 90.041% with lower variation and regret; W = 25 also has the highest mean among tested warm-ups.Neither increasing returns with K nor a unique optimum at K = 6 is supported; warm-up lengths use separate cosine schedules, and W = 25 remains fixed for the main results.Source: §4.3 · p. 7
  • Uniform proposals increase archive size and best-visited accuracy but lower selected accuracy and raise regret; removing mass redistribution also lowers selected accuracy, while three selectors perform comparably on average.Matching the uniform archive size offline does not recover the gap, and CMA-ES wins on all eight seeds; selector differences are small and not statistically significant in the reported cohort.Source: §4.3 · p. 7

4.4 DATASET TRANSFER

  • Without retuning, LESS obtains 69.615 ± 1.139% CIFAR-100 test accuracy and 43.720 ± 1.697% ImageNet16-120 test accuracy using the frozen configuration.The CIFAR-100 calibration sweep gives no clear evidence for selecting a different depth; on ImageNet16-120, mean selection regret is 2.662 percentage points despite every run visiting an architecture above the 99th percentile.Source: §4.4 · pp. 7–8

4.5 ACCURACY–COMPUTE SCALING

  • Across 30–120 search epochs, mean validation accuracy rises from 89.976% at 257 seconds to 90.203% at 567 seconds, then is 90.183% at 120 epochs.Selection regret falls from 0.679 to 0.525 points, while new genotypes per 30-epoch block decline from 95.5 to 7.9, explaining why additional compute changes the final selection less.Source: §4.5 · p. 8

4.6 DARTS-SPACE TRANSFER

  • With frozen NAS-Bench-201 search hyperparameters, LESS obtains 96.95 ± 0.14% CIFAR-10 and 82.43 ± 0.80% CIFAR-100 test accuracy in the DARTS space.Each search takes approximately 43.5 minutes on one RTX 4090; the results are close to established DARTS-space methods without DARTS-specific hyperparameter tuning.Source: §4.6 · p. 8, Table 2 · p. 9

5 DISCUSSION

LESS’s calibration advantage appears chiefly in choosing among visited architectures rather than finding better ones. Its search strategies trade off discovering strong genotypes against reliably returning them, and calibration also shifts the structure of selected cells.

5 DISCUSSION

  • Selected validation accuracy rises by 0.577 percentage points while best-visited accuracy changes by only 0.054 points, indicating lower selection regret.Alternative selectors perform similarly on recorded trajectories, so the gain cannot be attributed to Borda alone.Source: §5 · p. 9
  • Uniform proposals discover stronger architectures but return weaker ones than CMA-ES under the same evaluator; CMA-ES discovers few new genotypes late in search.The discussion suggests renewing exploration while preserving reliable candidate comparisons.Source: §5 · p. 9
  • Calibration is associated with more convolutional edges in selected NB201 cells, alongside fewer skip connections and null edges.These descriptive differences do not establish which structural changes cause the accuracy improvement.Source: §5 · p. 9
  • Transfer shows the frozen configuration is usable across three NB201 datasets and the larger DARTS space, but does not establish universal ranking reliability.CIFAR-100 has weaker proxy–truth agreement, and DARTS uses only three searched architectures per dataset.Source: §5 · p. 9

6 CONCLUSION

LESS combines brief fair supernet training, candidate-conditioned hard-path updates, and evolutionary search. It completes NAS-Bench-201/CIFAR-10 search in 409 seconds and transfers without search-hyperparameter retuning to other datasets and the DARTS space.

6 CONCLUSION

  • 409 seconds suffices for NAS-Bench-201/CIFAR-10 search, and the frozen search configuration transfers without retuning to other datasets and the DARTS space.The conclusion attributes the improved returned architecture to a few balanced, task-dependent updates, even when the best visited architecture changes little.Source: §6 · p. 9
  • LESS combines brief fair supernet training, candidate-conditioned hard-path updates, and evolutionary search.The work uses public image-classification benchmarks and no human subjects or private data.Source: ETHICS STATEMENT · p. 9
  • The released implementation records configurations, data schedules, proposal streams, candidate scores, winner states, timing boundaries, and artifact checksums.A reproduction gate found exact agreement for all 780 candidate records, including proxy values and Borda selection.Source: REPRODUCIBILITY STATEMENT · p. 10

A EXPERIMENTAL SCOPE, COHORTS, AND REPORTING CONVENTIONS

This appendix specifies the evidence cohorts, frozen implementation, and reporting conventions used to evaluate LESS. It distinguishes independent search replication from component and transfer analyses, and defines how candidate selection and benchmark outcomes are measured.

A EXPERIMENTAL SCOPE, COHORTS, AND REPORTING CONVENTIONS

  • The principal NB201 comparison pools an initial 16-seed cohort and an eight-seed prospective replication; component and transfer analyses reuse a fixed eight-seed subset rather than adding independent headline replications.For NB201, benchmark validation accuracy is the primary endpoint, test accuracy is descriptive, and search minibatch loss is a different quantity; DARTS reports highest test accuracy during 600-epoch retraining separately from final-epoch accuracy.Source: A EXPERIMENTAL SCOPE, COHORTS, AND… · p. 13, §A.1 · p. 13

B.1 FROZEN HYPERPARAMETERS

  • The frozen configuration specifies calibration, optimizer, diversity, initialization, and network settings, with NB201 and DARTS using 16 initial channels and eight cells.CMA scale is 0.30, distinct from the 0.45 standard deviation used to perturb the initial latent mean; benchmark accuracy does not enter initialization.Source: Table 4 · p. 13, §B.1 · p. 14

B.2 SPLITS AND TRANSFORMS

  • CIFAR search divides its official training examples into 25,000 warm-up/calibration/BN-refresh examples and 25,000 architecture-validation examples; ImageNet16-120 uses its canonical training split and NB201 x-valid split.Warm-up horizons use distinct cosine schedules rather than checkpoints from one longer trajectory.Source: §B.2 · pp. 14–15

B.3 PURPOSE-KEYED DATA SCHEDULES

  • Purpose-keyed data schedules keep calibration, BN refresh, and scoring streams deterministic and independent, while extending calibration does not shift BN or scoring data.At K = 6, each candidate uses six calibration batches, eight BN batches, and eight scoring batches; the calibration examples are shared across siblings and paths.Source: §B.3 · p. 15

C.1 NB201 LATENT VARIABLES AND HARD EXECUTION

  • NB201 decodes a 30-dimensional latent representation into six hard operation choices, with probability mass redistribution acting as a decoding prior rather than a hard-path weighting.The 15,625-cell space uses five operations per edge; architecture variables are not optimized by backpropagation during warm-up or calibration.Source: §C.1 · p. 15

C.2 BALANCED NB201 GRADIENTS AND STATE SEMANTICS

  • Each NB201 calibration update places the candidate first in a five-path group that executes every operation once per edge, averages the paths’ gradients, and steps weights after all backward passes.Candidate weights and BN buffers are distinguished; inheritance retains calibrated weights and pre-refresh buffers, while a separate BN refresh state is used for scoring.Source: §C.2 · pp. 15–16

C.3 CMA UPDATES, INCUMBENT STATE, AND BOUNDED DEDUPLICATION

  • The incumbent and CMA distribution mean remain distinct; accepted proposals are deduplicated, with genuine previously drawn duplicates used to fill slots only after the draw cap.The final selector accumulates normalized Borda reward over genotype loss rankings and selects the genotype with the maximum score, applying stated tie-breaks.Source: §C.3 · p. 16, §C.4 · p. 16

C.5 COMPUTATIONAL ACCOUNTING

  • Default search workload excludes warm-up and final retraining; paired differences are aggregated by seed, and bootstrap intervals, sign tests, and benchmark endpoints follow specified conventions.Benchmark joins use 200-epoch values; proxy–truth correlations and other retrospective diagnostics are defined over distinct visited genotypes.Source: §C.5 · p. 16, Table 7 · p. 16, D STATISTICAL DEFINITIONS AND… · p. 17

E CIFAR-10: PRINCIPAL PAIRED EVALUATION

The principal CIFAR-10 evaluation combines 24 paired searches, with component controls and operation-composition analyses providing additional descriptive evidence. The reported calibration advantage is primarily in the architecture returned rather than the best architecture visited.

E CIFAR-10: PRINCIPAL PAIRED EVALUATION

  • Calibration improves selected validation accuracy by 0.577 percentage points across the pooled cohorts, while best-visited accuracy changes by about 0.054 points and selection regret falls by about 0.523 points.The pooled gain combines separately reported cohorts descriptively; the decomposition does not isolate a single causal component.Source: Table 8 · p. 17, Table 9 · p. 18, E CIFAR-10: PRINCIPAL PAIRED EVALUATION · p. 18

E.1 OPERATION COMPOSITION

  • Returned-cell operation counts describe outcomes of the frozen decoder and search, not additional constraints on the architectures.Table 12 covers mean selected-cell operation counts across the six edges of each cell.Source: Table 12 · p. 19, §E.1 · p. 19

F COMPONENT CONTROLS AND ALTERNATIVE SELECTION RULES

  • The component-control tables report matched controls on S8 and per-seed selected validation, with test accuracy designated report-only.The supplied passages identify the tables’ scope but do not provide their cell values.Source: Table 13 · p. 19, Table 14 · p. 19

F.1 MORE EVALUATIONS WITHOUT CALIBRATION

The controls examine how evaluation budget, proposal strategy, candidate conditioning, calibration depth, warm-up, and search duration relate to selection quality and visited-pool quality.

F.1 MORE EVALUATIONS WITHOUT CALIBRATION

  • The expanded K0 control used 2,340 evaluations versus 780 for K6 and cost 529.0 versus 410.2 seconds; its 0.643-point validation difference had an interval spanning zero.This was an expanded-budget comparison, not an equal-time experiment; the K0 run cost about 1.29× as much.Source: §F.1 · p. 19

F.2 CMA VERSUS UNIFORM PROPOSALS AND THE DECODER PRIOR

  • CMA exceeded uniform on all eight matched seeds by 1.030 validation points, although uniform’s larger visited archive reached better best-visited validation accuracy.The selected-validation effect of keeping the decoder prior was 0.548 points, with an interval spanning zero.Source: §F.2 · p. 19

F.3 REPLAY OF THE FINAL SELECTION RULE

  • The offline selector replay on 16 K6 trajectories found that close alternatives do not establish a Borda-specific gain and required no additional candidate evaluations.The table’s “final incumbent” is the order-zero candidate entering the final epoch, not the incumbent after the last update.Source: Table 15 · p. 20, §F.3 · p. 20

G CANDIDATE-CONDITIONED VERSUS CANDIDATE-INDEPENDENT CALIBRATION

  • Conditioned calibration had a mean 1.009-point validation advantage over independent calibration, but the median was 0.057 and the evidence did not establish a uniform per-seed benefit.There were five wins, two ties, and one loss; the t interval included zero, while the percentile bootstrap interval was positive.Source: Table 18 · p. 20, G CANDIDATE-CONDITIONED VERSUS… · p. 20, Table 16 · p. 20

H.1 CALIBRATION DEPTH AT W25

  • At W25, all nonzero calibration depths had higher mean selected validation accuracy than K0, but the curve was nonmonotonic and did not establish a unique optimal depth.The frozen choice remained K6; greater pool-wide rank correlation at larger K did not guarantee a better returned architecture.Source: Table 19 · p. 21, §H.1 · p. 21

H.2 WARM-UP HORIZON AT K6

  • Shortening warm-up to five epochs increased seed variability and selection regret, while extending it beyond 25 epochs added cost without a resolved mean improvement.Each warm-up length used its own cosine schedule, so these comparisons change both training duration and learning-rate path.Source: Table 22 · p. 21, §H.2 · p. 22

I LONGER SEARCH AND LATE-TRAJECTORY CONCENTRATION

J UNIFORM-ARCHIVE SIZE AND REPEATED-EVALUATION ANALYSES

The archive analyses qualify comparisons between CMA and uniform search, while the transfer study reports CIFAR-100 outcomes for the frozen K6 configuration and limits what the depth curve establishes.

J UNIFORM-ARCHIVE SIZE AND REPEATED-EVALUATION ANALYSES

  • Uniform’s full archive had higher global proxy-truth correlation despite worse selected accuracy; reducing archive size did not restore CMA’s result, and the controls cannot isolate archive-size effects.CMA averaged 6.38 epoch appearances per genotype versus about 1.11 for full uniform, while the replay protocols also changed other aspects of the trajectories.Source: Table 29 · p. 23, J UNIFORM-ARCHIVE SIZE AND… · p. 23

J.1 REPEATED MEASUREMENTS

  • Across 245 genotypes evaluated in multiple epochs, mean stored loss improved Spearman correlation by 0.029, but both intervals included zero; incumbent-involved observations were not independent repeats.Chance redraws showed a more consistent secondary gain of about 0.055; the retrospective findings motivate, but do not replace, a prospective equal-evaluation experiment.Source: Table 30 · p. 24, §J.1 · p. 24

K CIFAR-100: FROZEN TRANSFER AND CALIBRATION DEPTH

L IMAGENET16-120 AND MATCHED DATASET DIAGNOSTICS

The ImageNet16-120 results separate within-pool selection from exploration; the accompanying sections document returned NAS-Bench-201 cells and the topology-aware DARTS adapter.

L IMAGENET16-120 AND MATCHED DATASET DIAGNOSTICS

  • ImageNet16-120 report-only test accuracy was 43.720 ± 1.697%, and roughly 75% of aggregate normalized regret came from selection within the visited pool.Cross-dataset comparisons use percentiles, normalized regret, pool coverage, and ranking correlations rather than raw accuracy differences; ImageNet16-120 component search time was 453.1 seconds.Source: Table 34 · p. 25, Table 36 · p. 25, L IMAGENET16-120 AND MATCHED DATASET… · p. 26

M RETURNED NAS-BENCH-201 ARCHITECTURES

N DARTS SEARCH-SPACE ADAPTER

  • The DARTS adapter decodes 224 latent logits into hard cells by retaining two distinct sources per node and each retained edge’s best non-null operation.The none operation influences redistribution but is never an active DARTS edge.Source: N DARTS SEARCH-SPACE ADAPTER · p. 27

N.1 TOPOLOGY-AWARE BALANCE

  • Because DARTS paths activate only two sources per node, the adapter uses approximate conditional operation balance; at K6, each update contains the focal path plus six valid complements.That gives seven paths and 42 passes; the adapter does not claim every eligible operation executes on every candidate edge in each update.Source: §N.1 · pp. 27–28

O DARTS RETRAINING AND COMPLETE PER-SEED RESULTS

The DARTS section specifies how selected cells were retrained and reports search dynamics and cell encodings; the supplied passages do not give the outcome values.

O DARTS RETRAINING AND COMPLETE PER-SEED RESULTS

  • Selected cells were retrained from scratch on all 50,000 CIFAR training images, with maximum test accuracy over 600 epochs as the primary endpoint and final-epoch accuracy supplementary.Each architecture was retrained once with seed 123, so the reported standard deviation describes variation across searched architectures under that fixed retraining seed.Source: O DARTS RETRAINING AND COMPLETE… · pp. 28–29, Table 40 · p. 28

O.1 SEARCH DYNAMICS AND MEMORY

  • All six selected searches ran 60 epochs and 780 candidate events without fallback; architecture seed 184271 returned the same genotype in separate CIFAR-10 and CIFAR-100 searches.The repeated selection should not be counted as two distinct cell pairs.Source: §O.1 · p. 29, Table 43 · p. 29

O.2 NORMAL AND REDUCTION CELLS

  • The normal and reduction cell listings encode each intermediate-node edge as an operation-source pair and concatenate nodes [2, 3, 4, 5].The listed cells cover every reported DARTS search.Source: §O.2 · p. 29, Table 44 · p. 29

P RUNTIME BOUNDARIES, HARDWARE, AND REPRODUCIBILITY

This appendix separates runtime accounting and reproducibility checks from earlier development cohorts and exploratory CMA-scale studies. It reports the boundaries and qualifications needed to interpret those records and results.

P RUNTIME BOUNDARIES, HARDWARE, AND REPRODUCIBILITY

  • Runtime components do not exhaust process time; NB201 and DARTS use distinct timing boundaries, and recorded hardware and determinism settings support reproducibility.NB201 component sums exclude startup and dataset initialization included in end-to-end manifests; DARTS active runtime excludes setup, resume downtime, and retraining.Source: P RUNTIME BOUNDARIES, HARDWARE, AND… · p. 30, Table 45 · p. 30

P.1 VERIFICATION RECORDS AND IMPLEMENTATION EQUIVALENCE

  • Verification records preserve configurations and candidate outputs, while replay and completion checks test selector reproducibility and implementation equivalence.The E120 gate matched 2,170 raw proposals and 780 ordered candidates; DARTS checkpoints support validated recovery without counting abandoned-attempt time as active runtime.Source: §P.1 · pp. 30–31

Q EARLIER DEVELOPMENT EVIDENCE KEPT SEPARATE

  • Earlier development evidence is kept separate from the principal synthesis, with historical labels subordinated to seed lists, schedule generations, and code revisions.The appendix does not use proxy-pilot analyses.Source: Q EARLIER DEVELOPMENT EVIDENCE KEPT… · p. 31

Q.1 EARLIER 16-PAIR NB201 COHORT

  • The earlier 16-pair cohort had a pooled validation gain of 0.517 points, but remains developmental support rather than part of the newer 24-seed synthesis.The cohort had 12 wins and four losses; its t interval was [0.008, 1.026] and bootstrap interval [0.058, 0.967].Source: §Q.1 · p. 31, Table 47 · p. 31

Q.2 EARLIER CIFAR-100 TRANSFER COHORT

  • The earlier CIFAR-100 transfer cohort used the original development seed list rather than S8, and its completed-run statistics remain distinct from the final depth-study estimate.One failed attempt was excluded and its completed replacement used the deterministic fallback once, at epoch 31.Source: §Q.2 · pp. 31–32, Table 48 · p. 31

Q.3 CMA-SCALE DEVELOPMENT AND ALTERNATIVE-SCALE STUDY

  • At σ = 0.45, the 16-seed K6-minus-K0 validation change was 0.454 points, with intervals spanning zero; this study remains separate from frozen σ = 0.30 evidence.The four-seed development grid favored 0.45 on validation mean, but its paired improvement over 0.30 was imprecise.Source: §Q.3 · p. 32, Table 50 · p. 32

R EVIDENCE PROVENANCE AND LIMITS OF INTERPRETATION

  • The principal limitations include small ablation and transfer seed counts, repeated S8 reuse, possible dependence across inherited-weight evaluations, and benchmark averages differing from fresh final training.DARTS balance is topology-constrained and approximate, and its three-architecture estimates use one final training seed.Source: R EVIDENCE PROVENANCE AND LIMITS OF… · p. 33

Related papers

Loading 2610.01468v1…