ImageNet Classification with Deep Convolutional Neural Networks
NeurIPS, 2012 · paper · AI and machine learning
Reports training a large deep convolutional neural network to classify 1.2 million ImageNet images into 1000 classes. It achieved top-5 test error of 17.0% on ILSVRC-2010, and a variant won ILSVRC-2012 with 15.3% versus 26.2% for the runner-up.
Checked: PDF from proceedings.neurips.cc, fetched 4 Oct 2026. Page numbers are those of that PDF.
2 findings
Substance Figure 2 caption gives the first-layer neuron count as "253,440" in "253,440–186,624–64,896–64,896–43,264–4096–4096–1000", but the figure's 55 × 55 maps with 48 + 48 kernels give 290,400.
- p. 5 Figure 2 caption
253,440–186,624–64,896–64,896–43,264–
- p. 5 Figure 2 drawing, first convolutional layer
55
- p. 4 Section 3.5, last paragraph
96 kernels of size 11×11×3
- p. 5 Figure 2 caption
253,440–186,624–64,896–64,896–43,264– 4096–4096–1000
Only the authors can say what is right here.
Judge's note: Read the Figure 2 drawing and caption on page 5 and the abstract on page 1. The drawing shows 55 × 55 maps with 48 kernels per GPU. The other caption counts match the drawing (27·27·256 = 186,624; 13·13·384 = 64,896; 13·13·256 = 43,264; input 224·224·3 = 150,528), but the first does not: 55·55·96 = 290,400. 253,440 equals 48·55·96, and 253,440/96 = 2,640 is not a perfect square, so no square map with 96 kernels gives it. The caption counts sum to 622,312; with 290,400 the sum is 659,272, nearer the abstract's "650,000 neurons". No explanation is given in the text. The likely intended value is 290,400, but this is tied to the 224/55 question (C007), so it is put to the authors.
Second check: another model agreed
Report a problem with this finding
Wording Introduction says of the convolutional layers "each of which contains no more than 1% of the model's parameters", but Section 3.5's "384 kernels of size 3 × 3 × 256" (conv 3) and "384 kernels of size 3 × 3 × 192" (conv 4), against "60 million parameters", give about 1.5% and 1.1%.
- p. 2 Section 1, contributions paragraph
each of which contains no more than 1% of the model’s parameters
- p. 5 Section 3.5 continuation
384 kernels of size 3 × 3 × 256
- p. 5 Section 4
60 million parameters
- p. 5 Section 3.5, text below Figure 2
The third convolutional layer has 384 kernels of size 3 × 3 × 256
Only the authors can say what is right here.
Judge's note: Read page 2 (contributions paragraph) and page 5 (Section 3.5 kernel sizes, Section 4 total). Recomputed weights per layer from the printed kernel counts and sizes: conv 3 is 884,736 weights, about 1.47% of 60 million; conv 4 is 663,552, about 1.11%; conv 1, 2 and 5 are under 1%. No statement in the paper redefines the 1% bound or the total. The claim is overstated for conv 3 and marginally for conv 4; whether "about 1%" was intended is for the authors.
Second check: another model agreed
Report a problem with this finding
4 minor (typos and the like; most checked by the judge only)
Reference or label "ILSVRC-2010 and ILSVRC-2012 competitions [2]" cites only reference [2], "Large scale visual recognition challenge 2010"; the 2012 challenge is reference [7] ("ILSVRC-2012, 2012"), not cited here.
- p. 2 Section 1, contributions paragraph
ILSVRC-2010 and ILSVRC-2012 competitions [2]
- p. 9 Reference [2]
Large scale visual recognition challenge 2010
- p. 9 Reference [2]
Large scale visual recognition challenge 2010.
Judge's note: Read page 2 and the reference list on page 9. [2] is Berg, Deng, Fei-Fei, "Large scale visual recognition challenge 2010. www.image-net.org/challenges. 2010"; [7] is the ILSVRC-2012 entry and is cited in Section 6. The URL in [2] is the general challenges page, so [2] may be intended as a general pointer; this is a minor citation looseness.
Second check: another model agreed
Report a problem with this finding
Typo Abstract writes "ImageNet LSVRC-2010 contest" while the same abstract uses "ILSVRC-2012 competition" and Section 2 defines and uses "ILSVRC".
- p. 1 Abstract, second line
ImageNet LSVRC-2010 contest
- p. 2 Section 2
ILSVRC-2010 is the only version of ILSVRC
- p. 1 Abstract, later line
ILSVRC-2012 competition
Judge's note: Read pages 1 and 2. The abstract has "ImageNet LSVRC-2010" and, a few lines later, "ILSVRC-2012"; Section 2 defines ILSVRC. "ImageNet LSVRC" can be read as spelling out the I, so the meaning is unchanged; this is only a cosmetic inconsistency of abbreviation and either form could be legitimate.
Second check: none (judge only)
Report a problem with this finding
Typo Footnote 2 misspells "connected": "nor the fully-conneced layers which follow".
- p. 3 Footnote 2
nor the fully-conneced layers which follow
Judge's note: Checked on the page 3 image: footnote 2 prints "fully-conneced". It is in the PDF, not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.1 has a duplicated word: "while the kernels on on GPU 2 are largely color-specific".
- p. 7 Section 6.1, first paragraph
while the kernels on on GPU 2 are largely color-specific
Judge's note: Checked on the page 7 image: the line ends "while the kernels" and the next line begins "on on GPU 2". The duplication is in the published PDF.
Second check: none (judge only)