Very Deep Convolutional Networks for Large-Scale Image Recognition

Very Deep Convolutional Networks for Large-Scale Image Recognition Karen Simonyan Andrew Zisserman

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Very Deep Convolutional Networks for Large-Scale Image Recognition 2014 ICLR / arXiv 1409.1556 simonyan2014-vgg simonyan2014-vgg.txt, the project's staged copy 3483e747b87c494d94350db046b7471ec28bf9c5 arxiv-open artificial-intelligence

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them method Quotation withheld because this record is not cleared for redistribution. convolutional network depth in large-scale image recognition Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. mechanisms stack many very small (3x3) convolution filters to build very deep networks - depth (16-19 layers), not filter size, drives accuracy, while small filters keep parameters low and grow the effective receptive field Quotation withheld because this record is not cleared for redistribution. Accuracy stops improving once depth reaches 19 layers on this dataset, so the depth benefit is shown to saturate. what this limits: the depth claim on this dataset: THE ERROR RATE OF OUR ARCHITECTURE SATURATES WHEN THE DEPTH REACHES 19 LAYERS, BUT EVEN DEEPER MODELS MIGHT BE BENEFICIAL FOR LARGER DATASETS Quotation withheld because this record is not cleared for redistribution. The approach is computationally expensive: a single network takes weeks to train even on four high-end GPUs. what this limits: anyone planning to reproduce or extend the training: ON A SYSTEM EQUIPPED WITH FOUR NVIDIA TITAN BLACK GPUs, TRAINING A SINGLE NET TOOK 2-3 WEEKS Quotation withheld because this record is not cleared for redistribution. The depth result is licensed by a control: every other parameter was deliberately held fixed. what this limits: the claim that depth improves accuracy: the authors fixed the other architectural parameters and increased depth alone, so the result holds for the family of configurations that control defines Quotation withheld because this record is not cleared for redistribution. The configurations being compared were built to one recipe, which is what makes the comparison fair and also what bounds it. what this limits: a comparison between these networks and others: all the configurations compared here were designed using the same principles in order to isolate depth Quotation withheld because this record is not cleared for redistribution. The explanation offered for faster convergence is labelled a conjecture by the authors. what this limits: the claim that greater depth speeds convergence: the authors conjecture the nets required fewer epochs owing to implicit regularisation imposed by greater depth and smaller filters Quotation withheld because this record is not cleared for redistribution. multi-crop and dense evaluation run against each other, and in combination eliminates: an accuracy figure that belongs to one evaluation protocol - running both shows what the protocol contributes and what the network does · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. ConvNet ConvNet Quotation withheld because this record is not cleared for redistribution. weight layers weight layers Quotation withheld because this record is not cleared for redistribution. receptive field receptive field Quotation withheld because this record is not cleared for redistribution. implicit regularisation implicit regularisation Quotation withheld because this record is not cleared for redistribution. depth depth Quotation withheld because this record is not cleared for redistribution. depth - AS A COUNT OF WEIGHT LAYERS: the configurations differ only in this number, from 11 in network A to 19 in E, so 'depth' here is an integer in a controlled series and not a general property of an architecture Quotation withheld because this record is not cleared for redistribution. ConvNet - THE PAPER'S OWN WORD, AND THE ONLY ONE IT USES: it says ConvNet 44 times and never writes 'convolutional neural network' or 'deep learning', both of which this record is filed under; the expanded name and the field's later label are absent from the paper that made them famous Quotation withheld because this record is not cleared for redistribution. the conclusion - ABOUT REPRESENTATION DEPTH ON A CONVENTIONAL ARCHITECTURE: the paper's own summary is that depth is beneficial and that the state of the art can be reached with a conventional architecture substantially deepened, which is a narrower claim than 'deeper is better' Quotation withheld because this record is not cleared for redistribution. extends → prior convolutional network configurations Quotation withheld because this record is not cleared for redistribution. builds-on → Krizhevsky et al. (2012) ConvNet training procedure Quotation withheld because this record is not cleared for redistribution. extends → the classical ConvNet architecture of LeCun et al. (1989) Quotation withheld because this record is not cleared for redistribution. extends → Ciresan et al. (2011) small-filter networks Quotation withheld because this record is not cleared for redistribution. depth —operationalised-as→ weight layers depth weight layers strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. depth —causes→ implicit regularisation depth implicit regularisation strength of the relation: 0.6 — the record marks this weak Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says 'we conjecture' the faster convergence is due to implicit regularisation; hedged. Quotation withheld because this record is not cleared for redistribution. 3 × 3 — very small convolution filter (receptive field) used throughout the network Quotation withheld because this record is not cleared for redistribution. 16–19 — network depth that yields the improvement Quotation withheld because this record is not cleared for redistribution. 1000 — ILSVRC classification output dimension about this quote: QUOTE RECOVERED FROM THE PAPER'S OWN CHARACTERS. The recorded quote was LETTERS-ONLY IDENTICAL to this stretch of the staged text but not an exact substring of it - spaces, hyphens, ligature glyphs or mathematical operators differed, so it could never anchor. This is the file's own text between the same two letters: not produced by a rule about the damage, copied out of the file. Round-trip checked (the recovered span is letters-only equal to the original) and length-checked. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. Pushing ConvNet depth to 16-19 weight layers was the basis of an ImageNet Challenge 2014 submission that took first place in localisation and second in classification. about this quote: QUOTE REPAIRED: the extractor split a hyphen-space and the quote was recorded faithfully with the damage in it, which made it unanchorable. The repaired string is an EXACT SUBSTRING of the staged text - that is why it was accepted, not because a rule produced it. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. The learned deep representations generalise well, achieving state-of-the-art results on other image datasets. Quotation withheld because this record is not cleared for redistribution. convolutional network depth convolutional network depth Quotation withheld because this record is not cleared for redistribution. top-5 error top-5 error Quotation withheld because this record is not cleared for redistribution. receptive field receptive field Quotation withheld because this record is not cleared for redistribution. regression regression about this quote: QUOTE RECOVERED FROM THE PAPER'S OWN CHARACTERS. The recorded quote was LETTERS-ONLY IDENTICAL to this stretch of the staged text but not an exact substring of it - spaces, hyphens, ligature glyphs or mathematical operators differed, so it could never anchor. This is the file's own text between the same two letters: not produced by a rule about the damage, copied out of the file. Round-trip checked (the recovered span is letters-only equal to the original) and length-checked. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://arxiv.org/abs/1409.1556.