Learning Transferable Architectures for Scalable Image Recognition
2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018 · paper · AI and machine learning
Searches for a convolutional building block on the small CIFAR-10 dataset, then stacks it to build networks for ImageNet. The resulting NASNet models report 82.7% top-1 ImageNet accuracy and outperform human-designed models at several computational budgets.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
8 findings
Substance Abstract and Introduction say the model has "9 billion fewer FLOPS – a reduction of 28%" than the best human-invented architecture, but Table 2 gives DPN-131 32.0 B against NASNet-A (6 @ 4032) 23.8 B Mult-Adds, a difference of 8.2 B (25.6%).
- p. 1 Abstract
9 billion fewer FLOPS – a reduction of 28% in compu-tational demand from the previous state-of-the-art model
- p. 2 Introduction
while having 9 billion fewer FLOPS
- p. 7 Table 2
32.0 B
- p. 1 Abstract
9 billion fewer FLOPS – a reduction of 28% in compu-
Only the authors can say what is right here.
Judge's note: I checked the Table 2 image on p7. The 1.2% top-1 gain (82.7 vs 81.5) identifies DPN-131 as the comparison model. 32.0−23.8 = 8.2 B, which is 25.6%, not 9 B and 28%. No other row gives 9 B/28% (PolyNet gives 10.9 B, SENet 18.5 B). The paper uses no separate FLOPS figure that would explain the gap. Only the authors can say whether the text or the table is out of date.
Second check: another model agreed
Report a problem with this finding
Substance Section 4.1 says "For the task of image classification with CIFAR-10, we set N = 4 or 6", but Table 1 lists "NASNet-A (7 @ 2304)" and "NASNet-A (7 @ 2304) + cutout" (the 2.40% headline model).
- p. 5 Section 4.1
we set N = 4 or 6 (Figure 2)
- p. 6 Table 1
NASNet-A (7 @ 2304)
- p. 6 Table 1
NASNet-A (7 @ 2304) + cutout
Only the authors can say what is right here.
Judge's note: Section 4 (p4) defines the notation: in '4 @ 64' the first number is the number of cell repeats N. On that definition the Table 1 rows (p6) with 7 are outside the stated N = 4 or 6. Only the authors can say whether the text or the table label is wrong.
Second check: another model agreed
Report a problem with this finding
Reference or label Footnote 1 gives "800 GPUs for 28 days resulting in 22,400 GPU-hours" and "500 GPUs across 4 days resulting in 2,000 GPU-hours"; GPUs × days gives GPU-days, so the unit is wrong.
- p. 4 Footnote 1
used 800 GPUs for 28 days resulting in 22,400 GPU-hours
- p. 4 Footnote 1
500 GPUs across 4 days resulting in 2,000 GPU-hours
- p. 4 Footnote 1
GPUs for 28 days resulting in 22,400 GPU-hours
- p. 4 Footnote 1
uses 500 GPUs across 4 days resulting in 2,000 GPU-hours
Judge's note: Seen in the p4 image. 800×28 = 22,400 and 500×4 = 2,000, both in GPU-days. The '7×' claim is explicitly adjusted for hardware, so the ratio itself is not at issue.
Second check: another model agreed
Report a problem with this finding
Reference or label Table 1 labels the row 'DenseNet-BC (L = 100, k = 40) [26]' but gives its depth as 190; L denotes depth, so the label should evidently read L = 190.
- p. 6 Table 1, row 4
DenseNet-BC (L = 100, k = 40) [26]
- p. 6 Table 1
DenseNet-BC (L = 100, k = 40) [26] 190
Judge's note: Page 6 Table 1 image: the row reads 'DenseNet-BC (L = 100, k = 40) [26] | 190 | 25.6M | 3.46'. In the other DenseNet rows L equals the depth column (40, 100, 100), so L = 100 with depth 190 is inconsistent within the same row. The label is wrong; the data are not in question.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [8] ("Dual path networks. arXiv preprint arXiv:1707.01083, 2017.") has the same arXiv identifier as reference [70] (ShuffleNet, "arXiv:1707.01083, 2017").
- p. 9 Reference [8]
Dual path networks. arXiv preprint arXiv:1707.01083, 2017.
- p. 10 Reference [70]
devices. arXiv preprint arXiv:1707.01083, 2017.
- p. 9 Reference [8]
path networks. arXiv preprint arXiv:1707.01083, 2017.
- p. 10 Reference [70]
arXiv preprint arXiv:1707.01083, 2017.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that 1707.01083 is ShuffleNet's identifier and DPN's is 1707.01629. Independent of that, the page images of p9 and p10 show two distinct papers, [8] and [70], with the identical identifier 1707.01083, so at least one entry is wrong.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [17] (Fukushima) gives a malformed page number: "Biological Cybernetics, page 93202, 1980."
- p. 9 Reference [17]
Biological Cybernetics, page 93202, 1980.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the Neocognitron paper occupies pp. 193–202. The p9 page image shows "page 93202", which is not a plausible page number in any case, so the entry is garbled.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [29] (Ioffe & Szegedy, Batch normalization) gives the venue as "In International Conference on Learning Representations, 2015."
- p. 9 Reference [29]
In International Conference on Learning Representations, 2015.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that Batch Normalization appeared at ICML 2015, not ICLR. The p9 page image shows ICLR 2015 as printed.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [30] (Jozefowicz et al., "An empirical exploration of recurrent network architectures") gives the venue as "International Conference on Learning Representations, 2015".
- p. 9 Reference [30]
exploration of recurrent network architectures. In Interna-
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that this paper appeared at ICML 2015. The p9 page image shows ICLR 2015 as printed.
Second check: another model agreed
Report a problem with this finding
28 minor (typos and the like; most checked by the judge only)
Reference or label Table 4 marks the ShuffleNet (2x) mini-val mAP as '24.5%†', but the Table 4 caption does not define †.
- p. 7 Table 4, ShuffleNet (2x) row and caption
24.5%†
Only the authors can say what is right here.
Judge's note: Page 7 image: the Table 4 caption has no † note. The only † definition is in the Table 3 caption, which is about top-1 accuracy taken from an open-source implementation. Only the authors can say what the † on this mAP means.
Second check: another model agreed
Report a problem with this finding
Wording Introduction says "the smallest version of NASNet achieves 74.0% top-1", but Table 3 lists NASNet-B (488M Mult-Adds) and NASNet-C (4.9M params), both smaller than NASNet-A (4 @ 1056) at 5.3M / 564M.
- p. 2 Introduction, para 2
the smallest version of NASNet achieves 74.0% top-1 ac-
- p. 7 Table 3
NASNet-C (3 @ 960)
Judge's note: Checked Table 3 on p7. The surrounding sentence is about varying cells and filters within the NASNet-A family, so 'smallest NASNet-A' is probably what is meant. The wording is loose rather than wrong. The abstract says 'a small version'.
Second check: another model agreed
Report a problem with this finding
Wording Conclusion states "In all cases, the accuracy of the resulting model exceeds all human-designed models", but in Table 2 SENet [25] and NASNet-A (6 @ 4032) both show 82.7 top-1 / 96.2 top-5, a tie rather than an improvement.
- p. 8 Conclusion, para 1
all cases, the accuracy of the resulting model exceeds all human-designed models
- p. 2 Introduction, para 2
accuracies superior to all human-invented models at equiv-
- p. 7 Table 2, SENet row
82.7
- p. 8 Conclusion
the accuracy of the resulting model exceeds all
Only the authors can say what is right here.
Judge's note: Checked Table 2 on p7 and the Conclusion on p8. The top rows of Table 2 tie (82.7/96.2) and the NASNet model uses fewer Mult-Adds (23.8B vs 42.3B). The Introduction sentence on p2 is qualified with "at equivalent or smaller computational budgets", and under that reading it holds. The comparisons the reviewer raised across blocks (80.8 vs 80.9 for ResNeXt-101, 78.6 vs 78.8 for Inception V3) set models of very different compute against each other, so they do not undercut the qualified claim. The unqualified Conclusion sentence does overstate the tie, and a later paragraph of the same Conclusion says only "approach or exceed state-of-the-art". Wording issue only.
Second check: another model agreed
Report a problem with this finding
Wording The example constructed block in Figure 3 uses a "2 x 2 maxpool", but the operation set in Section 3 lists only 3x3, 5x5 and 7x7 max pooling.
- p. 4 Figure 3, right panel
2 x 2 maxpool
- p. 3 Section 3 operation list
3x3 max pooling
Judge's note: Checked the Figure 3 image (p4) and the operation list (p3). The figure is illustrative, but the operation it shows could not be selected from the stated search space.
Second check: another model agreed
Report a problem with this finding
Wording Section 4 says "the pool of workers in the workqueue consisted of 500 GPUs" (and footnote 1 uses 500 GPUs), but Appendix A.3 says "we use a child worker pool size of 450, which means there are 450 networks being trained on 450 GPUs".
- p. 4 Section 4, para 1
the pool of workers in the workqueue consisted of 500 GPUs
- p. 11 Appendix A.3
we use a child worker pool size of 450, which means there are 450 networks being trained on 450 GPUs
- p. 4 Section 4, first paragraph
500 GPUs.
- p. 11 Appendix A.3
which means there are 450 networks being trained on 450 GPUs concurrently at any time
Only the authors can say what is right here.
Judge's note: Both passages checked on the images (p4, p11). Both describe the child-worker pool of the workqueue. The paper never says that the other 50 GPUs were used for the controller or for overhead.
Second check: another model agreed
Report a problem with this finding
Wording Section 3 says "random search is slightly worse than reinforcement learning", while Section 4.4 says "RL is significantly better than the best model found by RS by over 1%".
- p. 4 Section 3, last paragraph
we find that random search is slightly worse than reinforcement learning
- p. 8 Section 4.4
RL is significantly better than the best model found by RS by over 1%
- p. 4 Section 3
random search is slightly worse than reinforcement learning
- p. 8 Section 4.4
significantly better than the best model found by RS
Judge's note: Both passages checked (p4, p8). The same comparison is described as 'slightly' in one place and 'significantly' in the other. It is a framing inconsistency, not a numerical one.
Second check: another model agreed
Report a problem with this finding
Wording The Table 1 caption says 'All results for NASNet are the mean accuracy across 5 runs.' but the table reports 'error rate (%)'.
- p. 6 Table 1 caption
All results for NASNet are the mean accuracy across 5 runs.
Judge's note: Page 6 image: the only metric column is headed 'error rate (%)', and the NASNet values are error rates (e.g. 2.40). The caption misnames the metric. This is a wording issue.
Second check: another model agreed
Report a problem with this finding
Wording The text says 'our resulting network operating on images of the same spatial resolution (800 × 800) achieves mAP = 40.7%, exceeding equivalent ... (i.e. Inception-ResNet-v2)', but Table 4 lists the Inception-ResNet-v2 systems at 600 × 600 and 600 × 1000.
- p. 8 Section 4.3, first lines
network operating on images of the same spatial resolution (800 × 800) achieves mAP = 40.7%
- p. 7 Table 4
600 × 1000
- p. 8 Section 4.3, top of left column
(800 × 800) achieves mAP = 40.7%, exceeding equivalent
Only the authors can say what is right here.
Judge's note: Page 8 text and page 7 Table 4 images. The only 800-resolution baseline is ResNet-101-FPN at '800 (short side)'. The two Inception-ResNet-v2 rows that the sentence compares against are 600×600 (G-RMI) and 600×1000 (TDM). The sentence does not say which system shares the 'same spatial resolution'. The 4.0% margin is handled in C027.
Second check: another model agreed
Report a problem with this finding
Wording 'achieves mAP = 40.7%, exceeding ... Inception-ResNet-v2 by 4.0%': on test-dev, Table 4 gives a gap of 3.9 (vs TDM 36.8%) or 5.1 (vs G-RMI 35.6%); 4.0 matches only the mini-val gap (41.3 vs 37.3).
- p. 8 Section 4.3
achieves mAP = 40.7%, exceeding equivalent object detection systems based off lesser performing image featurization (i.e. Inception-ResNet-v2) by 4.0%
- p. 7 Table 4
36.8%
Only the authors can say what is right here.
Judge's note: Page 7 Table 4 and page 8 text images. 40.7 is the test-dev value, yet the 4.0 margin matches only the mini-val numbers. The sentence does not say which split or which baseline the 4.0 comes from.
Second check: another model agreed
Report a problem with this finding
Wording Section 4.3 states "achieves mAP = 40.7%, exceeding ... (i.e. Inception-ResNet-v2) by 4.0%", but Table 4 shows test-dev 40.7 vs 36.8 (TDM), a gap of 3.9. The 4.0 gap appears only on mini-val (41.3 vs 37.3).
- p. 8 Section 4.3
Inception-ResNet-v2) by 4.0% [28, 52]
- p. 7 Table 4
36.8%
Judge's note: Checked Table 4 on p7 and the text on p8. The 4.0% matches the mini-val comparison with Inception-ResNet-v2 (TDM). Next to the quoted 40.7% test-dev figure, the test-dev gap is 3.9. The text does not say which split the margin uses. On resolution: "same spatial resolution (800 × 800)" plausibly refers to the 800-short-side ResNet-101-FPN row, while the Inception-ResNet-v2 baselines are at 600×600 and 600×1000. This is a minor ambiguity, not a material error.
Second check: another model agreed
Report a problem with this finding
Wording The Figure 6 x-axis 'Number of Models Sampled' runs to about 45,000 (RL) and 42,000 (RS), while Appendix A.3 says 'the search process is terminated after 20,000 child models have been trained'.
- p. 8 Figure 6 x-axis
Number of Models Sampled
- p. 11 Appendix A.3
the search process is terminated after 20,000 child models have been trained
Only the authors can say what is right here.
Judge's note: Page 8 Figure 6 image: the RL curves extend to about 45k and the RS curves to about 42k. Page 11 A.3 gives a 20,000-model budget. Neither Section 4.4 nor the caption says that the RL-vs-RS comparison used a different budget, or whether 'sampled' counts duplicates while 'trained' does not. The authors should clarify.
Second check: another model agreed
Report a problem with this finding
Wording Conclusion cites "face detection [50]", but reference [50] is "Facenet: A unified embedding for face recognition and clustering".
- p. 8 Conclusion, para 3
face detection [50]
- p. 10 Reference [50]
Facenet: A uni-fied embedding for face recognition and clustering
Judge's note: Checked the Conclusion on p8 and reference [50] on p10. The reference's own title describes face recognition and clustering, not detection, so the task is mislabelled. The general point about reusing ImageNet features is unaffected.
Second check: another model agreed
Report a problem with this finding
Typo Introduction misspells the word as "this transferrability by designing a search space"; it should be "transferability".
- p. 1 Introduction, para 2
this transferrability by designing a search space
- p. 1 Introduction
transferrability
Judge's note: Confirmed in the p1 crop. The abstract spells it 'transferability'.
Second check: none (judge only)
Report a problem with this finding
Typo 'surpassing the best previous best by over 4.0% [37]' repeats 'best'; also, Table 4 gives exactly 43.1 − 39.1 = 4.0, not 'over' 4.0, and the abstract says 'by 4.0%'.
- p. 8 Section 4.3
surpassing the best previous best by over 4.0% [37]
- p. 7 Table 4
39.1%
- p. 8 Section 4.3
best previous best by over 4.0% [37]
- p. 1 Abstract
surpass state-of-the-art by 4.0%
Judge's note: Page 8 and page 7 images. The repeated word is a plain copyedit slip. With the printed one-decimal values the gap is exactly 4.0. Unrounded values could exceed 4.0, so 'over' is at most a small wording inconsistency with the abstract.
Second check: none (judge only)
Report a problem with this finding
Typo Operation list reads "5x5 depthwise-seperable conv"; it should be "separable".
- p. 3 Section 3 operation list
5x5 depthwise-seperable conv
Judge's note: Seen in the p3 image. The 3x3 and 7x7 entries are spelled 'separable'.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 2 caption: "The choice for the number of times the Normal Cells that gets stacked between reduction cells" is ungrammatical.
- p. 3 Figure 2 caption
the number of times the Normal Cells that gets stacked
Judge's note: Seen in the p3 image. 'Cells that gets' has a number-agreement error and a stray 'that'.
Second check: none (judge only)
Report a problem with this finding
Typo Section 4 repeats a word: "For complete details of of the architecture learning algorithm".
- p. 4 Section 4, right column
For complete details of of the architecture
- p. 4 Sec 4
complete details of of the architecture
- p. 4
For complete details of of the architecture learning algo-
- p. 4
ctively. For complete details of of the architecture le
Judge's note: Seen in the p4 image.
Second check: code check
Report a problem with this finding
Typo Footnote 1 has a stray article: "Discounting the fact that the we use faster hardware".
- p. 4 Footnote 1
Discounting the fact that the we use faster hardware
Judge's note: Seen in the p4 image.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 4 and Figure 8 captions read "A single block is corresponds to two primitive operations"; the 'is' is superfluous.
- p. 5 Figure 4 caption
A single block is corresponds to two primitive operations
- p. 13 Figure 8 caption
A single block is corresponds
Judge's note: Seen in the p5 and p13 crops.
Second check: none (judge only)
Report a problem with this finding
Typo Misspelling: '224x224 images using ∼550M multliply-add operations' should read 'multiply-add'.
- p. 6 Section 4.2 continuation, left column
multliply-add operations
Judge's note: Seen on the page 6 crop (left column, start of the text below Figure 5). The text layer has the same spelling.
Second check: none (judge only)
Report a problem with this finding
Typo The Table 2 caption reads 'Note that the composite multiple-accumulate operations are calculated...'; the previous sentence says 'multiply-accumulate'.
- p. 7 Table 2 caption
composite multiple-accumulate operations
Judge's note: Page 7 image of the Table 2 caption: 'multiply-accumulate' appears first, then 'multiple-accumulate'. The second is a slip.
Second check: none (judge only)
Report a problem with this finding
Typo Garbled phrase 'by over 1% as measured by on CIFAR-10': either a word is missing after 'by' or 'by' is extra.
- p. 8 Section 4.4
by over 1% as measured by on CIFAR-10
- p. 8 Sec 4.4
as measured by on CIFAR-10
Judge's note: Seen on the page 8 image, right column, Section 4.4.
Second check: none (judge only)
Report a problem with this finding
Typo Section 4.3 has a redundant word: "surpassing the best previous best by over 4.0% [37]".
- p. 8 Section 4.3
best previous best by over 4.0%
Judge's note: The p8 page image shows "the best previous best". One "best" should be removed. As a side note, the margin over RetinaNet in Table 4 is exactly 4.0 (43.1 vs 39.1), not "over" 4.0.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [70] lists an author as "L. Mengxiao", which appears to swap given and family names (Mengxiao Lin, i.e. "M. Lin").
- p. 10 Reference [70]
X. Zhang, X. Zhou, L. Mengxiao, and J. Sun.
- p. 10 Reference [70]
X. Zhang, X. Zhou, L. Mengxiao, and J. Sun
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the co-author's family name is Lin and given name Mengxiao. The p10 image shows "X. Zhang, X. Zhou, L. Mengxiao, and J. Sun". The cited work is still identifiable, so this is minor.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix A.1 is ungrammatical: "All images are whitened and then undergone several data augmentation steps".
- p. 11 A.1
All images are whitened and then undergone
Judge's note: The crop of p11 shows the text as quoted. It should read "then undergo".
Second check: none (judge only)
Report a problem with this finding
Typo Appendix A.4 has a number disagreement: "Unlike [24], all depthwise separable convolution do not employ Batch Normalization".
- p. 11 A.4
all depthwise separable convolution do
Judge's note: The crop of p11 shows the text as quoted. It should read "convolutions".
Second check: none (judge only)
Report a problem with this finding
Typo Appendix A.7 has an extra word: "Evaluations are calculated using with a running average of parameters".
- p. 12 Appendix A.7
Evaluations are calculated using with a running average
Judge's note: The crop of p12 shows "using with". One of the two words should be removed.
Second check: none (judge only)
Report a problem with this finding
Typo The Figure 8 caption has the wrong word order: "As do we not concatenate the output hidden states".
- p. 13 Figure 8 caption
As do we not concatenate the output hidden states
Judge's note: The crop of p13 shows the caption as quoted. It should read "As we do not concatenate". The same caption also has "A single block is corresponds to".
Second check: none (judge only)