Deep Residual Learning for Image Recognition
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016 · paper · AI and machine learning
Presents a residual learning framework that eases training of much deeper neural networks by learning residual functions relative to layer inputs. Residual nets up to 152 layers reached 3.57% ImageNet test error, winning ILSVRC 2015 classification.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
0 findings
None that passed a second check.
8 minor (typos and the like; most checked by the judge only)
Wording "top-5 validation error of 4.49%. This single-model result outperforms all previous ensemble results (Table 5)" compares a validation-set error with ensemble errors that Table 5 reports on the test set.
- p. 7 Sec. 4.1, Comparisons with State-of-the-art Methods
top-5 validation error of 4.49%. This single-model result outperforms all previous ensemble results (Table 5).
- p. 6 Table 5 caption
The top-5 error is on the test set of ImageNet and reported by the test server.
Only the authors can say what is right here.
Judge's note: p.7 explicitly says 'validation error', and the Table 5 caption on p.6 says 'The top-5 error is on the test set of ImageNet'. Each set is labelled, but the sentence says 'outperforms' across the two sets without noting this. It is a wording point only. Could the authors confirm the intended comparison?
Second check: another model agreed
Report a problem with this finding
Wording "It has fewer parameters than other deep and thin networks such as FitNet [35] and Highway [42] (Table 6)": ResNet-110 has 1.7M parameters, but Table 6 also lists a 32-layer Highway net with 1.25M.
- p. 8 Sec. 4.2, left column, continuing from p.7
networks such as FitNet [35] and Highway [42] (Table 6)
- p. 7 Sec. 4.2, last paragraph
It has fewer parameters than other deep and thin
- p. 7 Table 6, Highway 32-layer row
1.25M
- p. 7 Table 6
Highway [42, 43] 32 1.25M 8.80
Only the authors can say what is right here.
Judge's note: Table 6 (p.7) shows FitNet at 2.5M, Highway-19 at 2.3M, Highway-32 at 1.25M and ResNet-110 at 1.7M. The claim holds for FitNet and Highway-19 but not for the Highway-32 row of the cited table. The authors probably meant the best-performing Highway, but the sentence does not say so.
Second check: another model agreed
Report a problem with this finding
Typo Sec. 4.1: "are the same as the above plain nets, expect that a shortcut connection is added" should read "except that".
- p. 5 Sec. 4.1, Residual Networks paragraph
are the same as the above plain nets, expect that a shortcut
Judge's note: The p.5 image shows 'expect' where 'except' is meant. This is a plain misspelling.
Second check: none (judge only)
Report a problem with this finding
Typo Ref. [43] reads "Training very deep networks. 1507.06228, 2015." with no 'arXiv:' prefix, unlike [42] and the other arXiv entries.
- p. 9 Reference [43]
networks. 1507.06228, 2015.
Judge's note: The p.9 image confirms that [43] gives the italic '1507.06228' with no repository name, whereas [42] has 'arXiv:1505.00387'. This is a cosmetic inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo The column header "areo" in Tables 10 and 11 should be "aero" (aeroplane).
- p. 11 Table 10 header
areo
Judge's note: The p.11 image shows 'areo' in the headers of both Table 10 and Table 11.
Second check: none (judge only)
Report a problem with this finding
Typo The Table 10 and 11 captions read "The system “baseline+++” include box refinement"; the verb should be "includes".
- p. 11 Table 10 caption
The system “baseline+++” include box refinement
- p. 11 Table 11 caption
The system “baseline+++” include
Judge's note: The p.11 image confirms 'include' in both captions. The subject 'The system' is singular, so this is a grammar slip.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix C: "are both in a per-class from, in contrast to [32]" should read "per-class form".
- p. 12 Appendix C, second paragraph
are both in a per-class from, in contrast to [32]
- p. 12 Appendix C
are both in a per-class from
Judge's note: The p.12 image shows 'from'. Two sentences earlier the same paragraph has 'designed in a per-class form'.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix C: "To avoid negative samples being dominate" and "contains a single dominate object" should use "dominant".
- p. 12 Appendix C, left column
avoid negative samples being dominate
- p. 12 Appendix C, right column
contains a single dominate object
- p. 12 C. ImageNet Localization
To avoid negative samples being dominate
- p. 12 Appendix C
negative samples being dominate
Judge's note: The p.12 image shows both instances. The verb 'dominate' is used where the adjective 'dominant' is needed.
Second check: none (judge only)