SuperVision wins the ImageNet Large Scale Visual Recognition Challenge

Filed under idea, and it is a moment. canon/proposals.md puts alexnet-2012 in the idea section, between samuel-checkers-1959 and sparse-moe-2017, and describes it as a paper: "ImageNet Classification with Deep CNNs (Krizhevsky, Sutskever & Hinton, 2012) — top-5 error from 26% to 15% on two gaming GPUs." The reclassification is not a quibble, because the reason idea fails is the same reason the entry is worth having. Almost nothing in this paper is an idea, and its authors say so on the page. The nonlinearity is credited to Nair and Hinton (2010) in the sentence that introduces it. The regularizer had been on arXiv for three months under five names. The architecture is described as "somewhat similar to that of the 'columnar' CNN employed by Cireşan et al." The convolutional network is LeCun's, from the 1980s. What the paper contributes, in its own words, is that they "trained one of the largest convolutional neural networks to date" and "achieved by far the best results ever reported on these datasets." That is a result, obtained on a date, in a scored public contest, announced in a room. Every clause of the house moment test is satisfied by it and almost none by the alternative.

The definition being applied is the one eliza-1966 argued out when it talked its way out of the kind, and which clippy-1996, siri-2011, expert-systems-collapse-1987, deep-blue-1997, watson-jeopardy-2011, alphago-move-37-2016, chatgpt-2022 and alphafold2-2020 have used since: a date on which something visibly happened in public. The submission deadline was Sunday 30 September 2012, 23:00 GMT — the challenge page says "There will be no further extensions" — and the results were presented at a workshop at ECCV 2012 in Florence on Friday 12 October 2012, with the full leaderboard released the following day. There is a date, a venue, a scored task with a held- back test set, an adjudicating body, six rival teams, and a printed table showing who came where. rosenblatt-perceptron-1958 settled the precedent explicitly: "programs filed as idea; deep-blue-1997 and chatgpt-2022, which are events with dates, are moment." This is an event with a date.

The stronger form of the argument is that filing it as idea gets the citation occasion backwards. proposals.md names the occasion correctly — "Cite when: compute-driven capability, GPU demand, why the substrate story is the real story" — and then files the entry under the kind that points at the architecture. If a reading cites this file for an architectural claim it is citing the wrong file; the architecture entry is lecun-lenet-1998, which does not exist. What happened in 2012 is that a set of known techniques, none of them new, was run on hardware that had not previously been pointed at them, and the result was not a marginal improvement but a rout. That is a fact about substrate and it belongs in the kind that records events.

The name is retroactive and the entry keeps it under protest. The word "AlexNet" does not appear in the paper's title, abstract, architecture section, results or discussion — I read those pages. The contest leaderboard calls the team SuperVision. The title field above uses the contest name and the id stays alexnet-2012 because that is what proposals.md calls it and ids are not the job's to renumber. A reading should be aware that the name it is using is one the field applied later, to the network rather than to the entry, and that it names one of three authors. Hinton's own summary, given to the Computer History Museum in 2025, allocates the credit differently: "Ilya thought we should do it, Alex made it work, and I got the Nobel Prize."

descends_from holds one id, and the descent is architectural rather than documented in citations. The paper does not cite Rosenblatt. It does not need to: the canon has already made this finding in its own pages, and it is rosenblatt-perceptron-1958 that makes it — "Everything that has since been called a neural network is a descendant of it" — while being precise about which part travels. What descends unbroken is "weighted sums, a threshold or other nonlinearity, a learned read-out layer, error-driven updates" and the vocabulary. What does not descend from Rosenblatt is the training method, which is Rumelhart, Hinton and Williams in 1986. That entry does not exist either. Using the sibling entry's own finding, rather than inferring a connection from subject matter, is what keeps this edge from being the kind of vibe-descent chatgpt-2022 warned about.

The proximate ancestors are four and not one of them is in canon/, which is a gap this entry names rather than invents around:

• The ImageNet dataset — Deng, Dong, Socher, Li, Li and Fei-Fei Li, CVPR 2009. Without it there is no contest and no result; the paper describes it as "over 15 million labeled high-resolution images belonging to roughly 22,000 categories," collected from the web and labeled through Mechanical Turk. It is not proposed under any id. It is the most conspicuous absence in the proposal list.
• LeNet and the convolutional network — LeCun's work from 1989 onward. Cited by the paper repeatedly. No id proposed.
• Backpropagation — Rumelhart, Hinton and Williams, 1986. proposals.md says outright that it was left out at the cap.
• DanNet — Cireşan, Meier, Masci, Gambardella and Schmidhuber at IDSIA, from 2011, which is the first deep GPU-trained CNN to win vision contests and which the paper cites. No id proposed. Section 4 grades what follows from that.

mcculloch-pitts-1943, which rosenblatt-perceptron-1958 says it would write first, is still the file that every neural entry in this canon will eventually want to descend from.

What it is

The contest, which is the part that makes this evidence

The ImageNet Large Scale Visual Recognition Challenge was run annually from 2010 by Jia Deng and Alex Berg with Sanjeev Satheesh, Hao Su, Aditya Khosla and Fei-Fei Li, out of Stanford, Stony Brook and MIT. The classification task in 2012 was fixed and unglamorous: 1,000 object categories, roughly 1.2 million labeled training images (1,281,167 by the organizers' later count), 50,000 validation images, and 150,000 test images whose labels were not released. Entrants submitted predictions; the organizers scored them. Two error rates were reported, top-1 and top-5, the latter being the fraction of test images for which the correct label is not among the model's five most probable.

The withheld test labels are what make this file usable. Nothing in the result depends on the winner's own account of the winner's own performance. The scoring was done by people who did not build the system, on data the builders could not see, against six other teams who had the same data and the same deadline. That property is rarer in 2026 than it was in 2012, and a reading reaching for this entry when a lab reports its own benchmark numbers should say so.

The competition, which was hand-designed features

The leaderboard for the 2012 classification task, in order:

| Team | Institution | Top-5 error | |---|---|---| | SuperVision | University of Toronto | 0.15315 — "Using extra training data from ImageNet Fall 2011 release" | | SuperVision | University of Toronto | 0.16422 — "Using only supplied training data" | | ISI | University of Tokyo | 0.26172 | | ISI | University of Tokyo | 0.26602 | | OXFORD_VGG | University of Oxford | 0.26979 | | XRCE/INRIA | Xerox/INRIA | 0.27058 | | University of Amsterdam | University of Amsterdam | 0.29576 |

Read the method descriptions and the shape of the field is unmistakable. ISI combined "Fisher Vectors" and "Graphical Gaussian Vectors" over "SIFT and CSIFT, but also LBP and GIST," fused by a passive-aggressive classifier. Oxford's VGG used "Fisher vector image representation and one-vs-rest linear SVM classifiers" over "SIFT and colour statistics." Xerox/INRIA used Fisher vectors over SIFT and color with product quantization and linear SVMs. Every entry below the winner is a hand-designed feature extractor feeding a linear classifier, and each of those features — SIFT, GIST, LBP — is a named piece of somebody's careful work from the preceding decade. SuperVision's description is one sentence about a network "trained on raw RGB pixel values."

The four teams behind the winner are separated from each other by less than four percentage points and from the winner by ten. That is the whole finding of the contest, and it is why the entry is a moment and not an incremental data point: the field's mature, well-tuned, mutually competitive best clustered tightly, and something from outside the cluster landed ten points clear of all of it. The same table shows SuperVision taking the top two places in the localization task as well, at 0.335463 and 0.341905.

watson-jeopardy-2011 already recorded the far side of this from eighteen months earlier: "the last moment when a hand-engineered ensemble could win a headline result." This is the table where that stopped being true.

The machine, which is two gaming cards in a bedroom

The paper's description of its own hardware is four sentences long and is the most consequential passage in it:

> In the end, the network's size is limited mainly by the amount of memory > available on current GPUs and by the amount of training time that we are > willing to tolerate. Our network takes between five and six days to train on > two GTX 580 3GB GPUs. All of our experiments suggest that our results can be > improved simply by waiting for faster GPUs and bigger datasets to become > available.

The GeForce GTX 580 was a consumer graphics card, launched 9 November 2010 at $499, 512 CUDA cores, 1.5 GB of GDDR5 in the reference design; the 3 GB version the paper used was a board-partner variant. It was, in 2012, a two-year-old part sold to people who wanted to play games. Two of them, plus a host machine, is the entire compute footprint of the result. The Computer History Museum's account, and IEEE Spectrum's, put the machine in Krizhevsky's bedroom at his parents' house.

The other numbers from the authors' own presentation deck are of the same scale: 27 GB of disk for the whole training set stored as JPEG, about 2 GB of RAM on each GPU, around 5 GB of system memory, written in Python/C++/CUDA, with loading, transforming, copying and training running as a four-stage pipeline so the GPUs never waited on the disk. Roughly 90 passes through the 1.2 million images.

Even the two-GPU split, which later became the visual signature of the architecture — the famous two-track diagram — was a memory workaround stated as such: "A single GTX 580 GPU has only 3GB of memory, which limits the maximum size of the networks that can be trained on it. It turns out that 1.2 million training examples are enough to train networks which are too big to fit on one GPU. Therefore we spread the net across two GPUs." Section 4 grades what became of that.

There is one more piece of hardware history worth carrying, because it is the counter-example to every retrospective account in which the chip vendor saw this coming. Hinton's telling, to Timothy B. Lee: after giving a talk at NIPS 2009 telling a thousand machine-learning researchers to buy Nvidia cards, "I sent an e-mail saying, 'Look, I just told a thousand machine-learning researchers they should go and buy Nvidia cards. Can you send me a free one?' … They said no." This is Hinton's recollection of his own email and is not otherwise documented here; it is carried as testimony, not as record.

The result, stated precisely

On ILSVRC-2010, where test labels were available and most of the paper's experiments were run: top-1 37.5%, top-5 17.0%, against 47.1%/28.2% for the best entry in the 2010 competition and 45.7%/25.7% for the best published result since.

On ILSVRC-2012, the contest itself, the paper reports a table rather than a number, and the distinction inside it matters:

| Model | Top-1 (val) | Top-5 (val) | Top-5 (test) | |---|---|---|---| | SIFT + FVs (second-best contest entry) | — | — | 26.2% | | 1 CNN | 40.7% | 18.2% | — | | 5 CNNs | 38.1% | 16.4% | 16.4% | | 1 CNN\ (pre-trained on ImageNet Fall 2011) | 39.0% | 16.6% | — | | 7 CNNs\ | 36.7% | 15.4% | 15.3% |

The famous 15.3% is an ensemble of seven networks, two of which were pre-trained on the whole 15-million-image ImageNet Fall 2011 release — extra data the other teams did not use, which is why the leaderboard files it separately. The best number obtained under the same data constraints as everyone else is 16.4%, five networks averaged, and a single network is 18.2%. The margin over the field is therefore 10.9 points stated one way and 9.8 points stated the more conservative way, and a single un-ensembled model trained on supplied data only is still eight points clear of the next team.

Every one of those framings is a rout. This entry gives all of them because a canon that grades overclaims should not itself quote the most flattering available number, and because "15.3 versus 26.2" is the form the comparison almost always takes in retellings.

The room in Florence

The results were presented on Friday 12 October 2012 at a workshop held alongside ECCV 2012 in Florence, run in conjunction with the PASCAL Visual Object Classes challenge workshop — which is a detail worth keeping, because the event now treated as the hinge of the field was a session inside somebody else's workshop. The surviving artifact is Krizhevsky's slide deck, still served from image-net.org, whose title slide carries the line "Paper with same name to appear in NIPS 2012."

The accounts of the room are consistent and secondhand: it was standing-room only, Yann LeCun and Fei-Fei Li were in it, and LeCun stood up afterwards and called the result "an unequivocal turning point in the history of computer vision. This is proof." That quotation is widely reported and is generally traced to Fei-Fei Li's 2023 memoir The Worlds I See. I did not read the memoir and could not reach a primary transcript; contemporaneous discussion on Google+ is cited by others but Google+ is dead and the Internet Archive is not reachable from this job. A reading quoting LeCun should attribute it as reported, not as recorded.

The paper itself appeared at NIPS 2012 in December, in Lake Tahoe, as Advances in Neural Information Processing Systems 25, pp. 1097–1105.

What was actually new, which is the part most retellings drop

The paper has a section ordering its own contributions "according to our estimation of their importance, with the most important first." Taken in that order, with the paper's own attributions:

• ReLU nonlinearity. Not theirs. "Following Nair and Hinton [20], we refer to neurons with this nonlinearity as Rectified Linear Units (ReLUs)" — Nair and Hinton, ICML 2010, Haifa. The paper adds: "We are not the first to consider alternatives to traditional neuron models in CNNs." Its contribution is the measurement that ReLUs reach 25% training error on CIFAR-10 six times faster than tanh, and the observation that "we would not have been able to experiment with such large neural networks for this work" without them.
• Training on multiple GPUs. An engineering response to a 3 GB memory ceiling, and by the paper's own account an architecture "somewhat similar to that of the 'columnar' CNN employed by Cireşan et al."
• Local response normalization. New here, and wrong; see section 4.
• Overlapping pooling. New here, and small: 0.4% and 0.3%.
• Dropout. Not theirs and not new to this paper: Improving neural networks by preventing co-adaptation of feature detectors, Hinton, Srivastava, Krizhevsky, Sutskever and Salakhutdinov, arXiv, 3 July 2012 — three months before the contest deadline, by an overlapping author list.
• Data augmentation. Translations, reflections and PCA colour jitter, standard practice.

So the honest inventory of novelty is: one normalization scheme that did not survive, one pooling change worth a third of a point, and a very well-engineered CUDA implementation of a known architecture, run at a scale nobody had run it at before, on a dataset that had only just come to exist. The code was public from the start — the paper footnotes code.google.com/p/cuda-convnet/ — which matters to how fast the result propagated and is a fact worth holding beside any 2026 argument about open weights.

Afterwards, in four movements

December 2012 — the auction. Hinton incorporated DNNresearch with Krizhevsky and Sutskever, a company with no product, and auctioned it during NIPS at Lake Tahoe. The bidders were Baidu, Google, Microsoft and the then- independent DeepMind; Baidu opened at $12 million. Hinton stopped the bidding and sold to Google. The price is reported as $44 million in Cade Metz's Genius Makers and most accounts derived from it; a first-hand participant account gives $42 million, and the two figures are not reconciled here. Google announced the acquisition in March 2013. Three people, one paper, four months.

2013 onward — the field switches. The organizers' own retrospective is blunt: "Following the success of the deep learning-based method in 2012, the vast majority of entries in 2013 used deep convolutional neural networks." The official ILSVRC-2013 classification winner was Clarifai at 0.11197, Matthew Zeiler's company, founded on the visualization work he and Rob Fergus published that year; Zeiler's academic entry under the team name ZF placed eighth at 0.13511. By 2014 GoogLeNet took it to 6.7%, by 2015 ResNet to 3.57%, and the challenge's final running in 2017 was won by SENet at 2.251%. Total elapsed time from "hand-designed features are the state of the art" to "the benchmark is saturated and retired": five years.

March 2025 — the code. The Computer History Museum released the 2012 source code under a BSD-2 license, after five years of negotiation begun in 2020, with the permission of Google, which had owned it since the DNNresearch sale. The release is the reason a 2026 reader can check what the thing actually was.

The people. Hinton shared the 2018 Turing Award with Bengio and LeCun and the 2024 Nobel Prize in Physics with Hopfield. Sutskever co-founded OpenAI in 2015, left in 2024, and founded Safe Superintelligence. Krizhevsky worked at Google on Photos and then the self-driving project and left in September 2017, having, in his own account to Quartz, lost interest in the work. The paper was ranked eighth on Nature's 2025 list of the 25 most-cited papers of the 21st century, with over 198,000 Google Scholar citations as of early 2025 — a count which by 2026 is larger and which nobody should quote to three significant figures.

Why a reading would cite it

The occasion is a compute claim, and this is the file that dates the claim's origin to a sentence rather than to a vibe.

The readings this project takes are saturated with capital expenditure. The 16 August 2026 midnight reading, to take the one immediately at hand, carries roughly $70 billion of off-balance-sheet chip financing, an Nvidia guarantee moving by $130 billion in a week, and a filed 13F in which two customers are 80.3% of the supplier's equity book. Every one of those numbers is downstream of a thesis: that buying more compute buys more capability, reliably enough to underwrite debt against. That thesis has a first published statement, it is one sentence long, it was made by three people with $1,000 of gaming hardware, and it is this:

> All of our experiments suggest that our results can be improved simply by > waiting for faster GPUs and bigger datasets to become available.

A reading that wants to say "the capex is a bet on scaling" can cite the bet's origin here. A reading that wants to say the bet has so far been paid is on firmer ground citing this than citing any vendor's roadmap, because the claim was made in advance, in public, in a paper whose result was adjudicated by strangers, and the fourteen years since are the out-of-sample test. The scaling-laws entry that would quantify this — scaling-laws-2020, Kaplan et al., corrected by Chinchilla — is proposed and unwritten. This file is the qualitative ancestor of that one, and the two should eventually be read together with the descent stated in whichever direction the later entry finds documented.

Second, and sharper: this entry is the correct scale-check on the substrate story, because it is the small end of it. The temptation, when a reading is looking at a $197.3 billion data-centre business, is to treat frontier capability as intrinsically capital-intensive — a thing only a state or a hyperscaler could ever have done. The record says otherwise about the beginning. Nvidia's revenue in the fiscal year that contained this result, FY2013 ended 27 January 2013, was $4.28 billion, and the company had no data-centre reporting segment at all. In FY2026 its data-centre revenue alone was $197.3 billion — roughly 46 times the entire company of the AlexNet year. The capability that started that curve was produced by two graduate students and their supervisor on hardware a teenager could buy, and the chip vendor had declined to give them one for free.

That comparison is the entry's most useful single act, and it cuts both ways, which is why it is honest. It is an AItopia-side fact — capability arrived from outside the incumbents, cheaply, and propagated in months because the code was public. It is equally an AImageddon-side fact — the same fourteen years turned a bedroom result into a concentration of ownership severe enough that a supplier's 13F now doubles as a market-structure disclosure. neuromancer-1984 holds the frame for the second reading and culture-banks-1987 for the first. A reading citing this entry for the concentration lens should be explicit that it is citing a trajectory and not an event: nothing about 2012 predicted who would own 2026.

Third: this is the canon's cleanest case of a capability that was demonstrated before it was claimed, which is the inverse of the pattern the readings mostly find. There was no announcement, no launch, no waitlist and no blog post. There was a deadline, a withheld test set, a table, and a talk in a room in Florence to an audience that had every professional incentive to be unimpressed. When a reading is grading a claimed-versus-demonstrated gap — the LENSES.md capabilities test, and the accelerant lens's requirement that a source show the method was actually used in the work — this file is the benchmark for what the demonstrated end of that gap looks like when it is real.

What this entry does not support. It is not evidence about any 2026 system, any 2026 model's capability, or any 2026 company's prospects. Nothing in it moves a needle. It is not an argument that scaling continues — the paper's sentence was a claim about 2012 hardware and 2012 datasets, it has been right for fourteen years, and "has been right for fourteen years" is a base rate, not a guarantee. A reading that cites this file as reassurance about the next purchase order is misusing it, and section 5 says how.

What it got right, and what it got wrong

Not required for a moment, but this entry is unusually rich in dated, falsifiable technical claims made by people who then had to live with them, and the grading is where most of its value sits.

Claim 1 — "our results can be improved simply by waiting for faster GPUs and bigger datasets." Made December 2012. Graded continuously. Right, and it is the most consequential sentence anyone in this canon has written about hardware.

The claim is in the introduction, immediately after the admission that the network's size was capped by 3 GB of memory. It is a claim about a bottleneck: not that the architecture was right, not that the ideas were finished, but that the binding constraint was the substrate and that the substrate was on a known improvement curve.

Fourteen years of scoreboard say the diagnosis was correct. The same benchmark went from 16.4% to 2.251% in five years, with the winning entries getting larger and the ideas changing much less than the size did. The pattern then repeated outside vision. Nvidia's data-centre business went from not being a reporting segment to $197.3 billion a year. The entire financing structure the 2026 readings keep finding — special-purpose vehicles borrowing against contracted chip cashflows — is a machine for converting this sentence into capital.

Two honest qualifications. First, the claim as written is about these results on this task, and it has been generalized far past its own scope by people who did not write it; that generalization is graded separately in section 5. Second, the sentence contains "and bigger datasets," which is the half that usually gets dropped in the retelling and the half that Chinchilla later showed had been systematically under-weighted. A reading quoting only the GPU clause is quoting three-quarters of a sentence.

Claim 2 — "the depth really is important." Made December 2012. Graded within two years. Right, and immediately outgrown.

The discussion states it twice: "It is notable that our network's performance degrades if a single convolutional layer is removed. For example, removing any of the middle layers results in a loss of about 2% for the top-1 performance of the network. So the depth really is important for achieving our results." The introduction adds the striking detail that the layers being removed "each of which contains no more than 1% of the model's parameters" — depth mattering independently of parameter count.

Correct, and the field took it much further much faster than the claim implied. Eight layers in 2012; VGG's finding in 2014 was that "a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16–19 weight layers"; ResNet in 2015 ran to 152. The claim is right and dated: eight layers was deep in 2012 and is a rounding error now. A reading that cites this entry for "depth is what matters" is citing a claim whose author would have revised it within eighteen months.

Claim 3 — local response normalization "aids generalization." Made December 2012. Refuted in print September 2014. Wrong.

This is the paper's one genuinely novel architectural component, and it is the one that died. The paper reports it carefully — "Response normalization reduces our top-1 and top-5 error rates by 1.4% and 1.2%, respectively", with the hyper-parameters k=2, n=5, α=10⁻⁴, β=0.75 chosen on a validation set, plus a confirming CIFAR-10 experiment at 13% without and 11% with.

Simonyan and Zisserman, arXiv, September 2014: "We note that none of our networks (except for one) contain Local Response Normalisation (LRN) normalisation: as will be shown in Sect. 4, such normalisation does not improve the performance on the ILSVRC dataset, but leads to increased memory consumption and computation time." LRN disappeared from the architectures that followed and has not come back.

This is the single most instructive claim in the entry and it deserves to be cited more than the famous ones. The measurement was honestly made and honestly reported by capable people, with a validation set and a replication on a second dataset, and it was still an artifact of one configuration rather than a property of neural networks. A 1.2-point improvement on one benchmark, obtained once, is not a finding. When a 2026 reading is looking at a lab's ablation table justifying an architectural choice, this is the precedent for how much weight that table can carry.

Claim 4 — the two-GPU column split. Made December 2012, as a memory workaround. Never a principle, and read as one for a decade.

The paper is explicit that the split exists because 3 GB was not enough, and reports it as worth 1.7% top-1 and 1.2% top-5 against a one-GPU net — while adding a footnote conceding that "this comparison is biased in favor of the one-GPU net, since it is bigger than 'half the size' of the two-GPU net." The paper also notes the split produced a real and reproducible phenomenon: "The kernels on GPU 1 are largely color-agnostic, while the kernels on GPU 2 are largely color-specific. This kind of specialization occurs during every run and is independent of any particular random weight initialization."

The finding is genuine; the design is not. The two-track diagram became the canonical image of the architecture and appears in a thousand lecture slides as though the split were part of the idea. It was a 3 GB memory ceiling. The lesson a reading can take from it: a great deal of what looks like architecture in a systems paper is the shape of that year's hardware, and it stops being copied the moment the hardware changes.

Claim 5 — dropout. Not this paper's claim, right at the time, largely retired since.

Dropout is credited by the paper to arXiv:1207.0580 (3 July 2012) and is applied "in the two globally-connected hidden layers" at p=0.5, with the paper noting it roughly doubles the iterations to convergence. It was the right answer to the problem the paper had: 60 million parameters, 1.2 million images, and, as the paper says, overfitting "a significant problem" even at that data scale. It was standard practice for years afterward.

It is not how the problem is solved now. Batch normalization, vastly larger datasets and different regularization regimes displaced it from the position it held in 2012. This is a claim that was correct and became obsolete without ever being wrong — a category the canon should keep distinct from claim 3, which was wrong on arrival.

Claim 6 — "many orders of magnitude to go." Made December 2012. Unresolved, and the least-quoted sentence in the paper.

> Thus far, our results have improved as we have made our network larger and > trained it longer but we still have many orders of magnitude to go in order to > match the infero-temporal pathway of the human visual system.

Fourteen years on, this is neither graded nor gradeable, because nobody agreed on a unit. Modern vision models exceed human top-5 accuracy on ImageNet's thousand categories by a wide margin and remain trivially defeatable in ways the human visual system is not. The sentence's honest status is open, and its value to this project is as a demonstration that the authors of the most triumphant result of their decade closed their paper with a statement of how far short of the target they were. A reading contrasting 2012's rhetoric with 2026's should have this quotation in hand.

What nobody claimed

No one in this record claimed general intelligence, imminent automation of anything, or that the technique would transfer beyond vision. The paper's most speculative sentence is a wish to try it on video. The claims that the result would generalize were made by other people, later, and were substantially correct — which is a different and rarer failure of forecasting than the ones this canon usually grades: the field's own most successful result was under-claimed by its authors, and the correct extrapolation came from outside them.

Commonly misused as

Not required for a moment, but this one has accumulated more folklore than any other entry in this canon except terminator-1984, and the folklore is load- bearing for arguments about 2026 spending.

"AlexNet invented deep learning" — or CNNs, or GPU training, or ReLU, or dropout.

It invented none of them, and the paper says so about four of the five. The convolutional network is LeCun's. ReLU is credited on the page to Nair and Hinton (2010). Dropout is a separate paper from three months earlier. And GPU- trained deep CNNs winning vision contests is Cireşan and Schmidhuber's group at IDSIA, which the paper cites: DanNet won four contests before ImageNet 2012 — 15 May 2011 (ICDAR, Chinese handwriting), 6 August 2011 (IJCNN, traffic signs), 1 March 2012 (ISBI, segmentation) and 10 September 2012 (ICPR, mitosis detection), the last of these twenty days before the ImageNet deadline. At IJCNN 2011 DanNet scored 0.56% error against a human average of 1.16%, which is a documented superhuman result in a controlled contest a full year before Florence.

Schmidhuber has pressed this publicly for years, holding that AlexNet was "similar to the earlier DanNet" and that ReLUs and dropout, while helpful, "are not really required to win vision contests," because "the only really important CNN-related task was to greatly accelerate known techniques for training CNNs through GPUs."

The grade this entry gives: Schmidhuber is right on the facts and the facts do not do the work he wants them to do. He is right that the techniques were not new, right that DanNet was first to win with a GPU CNN, right that the essential move was acceleration, and the AlexNet paper concedes the architectural similarity in print. What he is arguing against is a claim about priority; what made October 2012 a moment is not priority. It is that ImageNet was the benchmark the computer vision field had agreed to be judged on, that the entire mature competition showed up to it, and that the margin was ten points. DanNet's contests were real and its result was superhuman and the field did not turn. After Florence it turned within one year. The distinction the canon should hold: being first and being the event are different properties, and this entry is filed as moment precisely because it claims the second and not the first.

Both halves of that finding are usable. A reading tempted to write "X was the first to do Y" about a 2026 result should note how much work the word "first" is doing and check whether it is the claim that matters.

"AlexNet proved that compute is all that matters."

The paper does not say this and the record does not support it. The claim in the paper is that the binding constraint at that moment was memory and training time, and that relieving it would improve those results. It is a diagnosis of a bottleneck, not a theory of intelligence. The same paper reports that removing a layer holding under 1% of parameters degrades performance, that a nonlinearity choice is worth a 6× training speedup, and that its own novel normalization was worth 1.2 points — none of which are compute claims.

The generalized version, sometimes attached to Rich Sutton's 2019 "bitter lesson" — proposed as bitter-lesson-2019 and unwritten, and named as the missing entry by samuel-checkers-1959 — is a much stronger claim than anything in this paper, and a reading should not cite AlexNet for it. What this file supports is narrower and more useful: that in 2012, on this task, the field had been constrained by hardware without knowing it, and that the constraint was invisible until somebody removed it. Whether that is true of any given 2026 constraint is exactly the question, and this entry does not answer it.

"AlexNet beat humans."

It did not, and no ILSVRC entry did until 2015. The human figure everyone quotes — 5.1% top-5 error — comes from Russakovsky et al. and is the score of one trained annotator, Andrej Karpathy, on 1,500 test images after practising on 500; a second annotator who trained less scored around 12% on 280 images. The annotator's own caveat, published September 2014, is the part that should travel with the number: "Human accuracy is not a point. It lives on a tradeoff curve," with his estimate that dedicated experts might reach 3% and "about 2% being an optimistic error rate lower bound."

So the canonical "superhuman in 2015" milestone is a machine ensemble beating one motivated graduate student on a thousand-way labelling task involving 120 breeds of dog. That is a real and impressive engineering result and it is not a statement about vision. When a 2026 reading meets "exceeds human expert performance" in a benchmark claim, this is the entry that shows what the human number in such comparisons is usually made of — and it applies with particular force because the ImageNet version was carefully done and honestly caveated by the person who generated it. Most are not.

"Top-5 error went from 26% to 15%."

Nearly right, and it silently compares an ensemble of seven networks using extra training data against a competitor that used only the supplied data. Under matched conditions the winning number is 16.4%. Both comparisons show a rout; proposals.md uses the first; this entry gives both because the difference between them is exactly the kind of asterisk the readings are constantly having to find in vendor benchmark claims, and a canon that grades that behaviour should not model it.

"The deep learning revolution started on 30 September 2012."

30 September was a submission deadline. 12 October is the day it was shown to the field, 13 October the day the leaderboard went up, December the month of the paper and the auction, and 2013 the year the field actually switched — the organizers' own retrospective dates the switch to the composition of the 2013 entry list, not to 2012. If a reading needs one date, 12 October 2012 is the defensible one, on the same reasoning alphafold2-2020 applied to its own id: anchor to the adjudicated public event, not to the paperwork before it or the consequences after it.

The one this canon has to watch in itself

There is a standing pull, when a reading has a great deal of capital expenditure in front of it and needs a frame, to reach for a historical result that makes the spending look either obviously vindicated or obviously mad. This entry can be made to serve both and should serve neither. What it supports is a base rate and a shape: that the field's largest single discontinuity came from relieving a hardware constraint nobody had identified as binding, that it cost about a thousand dollars, that its authors under-claimed it, and that the fourteen years since have been a mostly-successful out-of-sample test of one sentence in its introduction. Everything past that is the reading's own judgment and should be labelled as such, per rule 2.

Sources

Primary, read directly:

• **Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton, ImageNet Classification with Deep Convolutional Neural Networks, Advances in Neural Information Processing Systems 25 (NIPS 2012), pp. 1097–1105.** Read as the camera-ready PDF from proceedings.neurips.cc. Every quotation in this entry of the abstract, the GPU-and-dataset sentence, the depth claim, the ReLU attribution to Nair and Hinton, the Jarrett remark, the two-GPU memory rationale and its footnote, the Cireşan "columnar" comparison, the LRN description and its hyper-parameters, the overlapping-pooling numbers, the 90-cycles/five-to-six-days training figure, Tables 1 and 2, and the whole of Section 7 Discussion is from this text, at pages 1–4 and 7–8. Note that the abstract listed on the NIPS proceedings index page differs from the camera-ready abstract — it describes 500,000 neurons and two globally-connected layers rather than 650,000 and three — so anything quoting "the abstract" should say which one; this entry quotes the paper.
• **Alex Krizhevsky, ImageNet Classification with Deep Convolutional Neural Networks, presentation slides, ILSVRC-2012 workshop.** Served from image-net.org/static_files/files/supervision.pdf. Source for the title slide's "Paper with same name to appear in NIPS 2012", the dropout slide, and the implementation figures (27 GB disk, ~2 GB per GPU, ~5 GB system memory, Python/C++/CUDA, four-stage pipeline).
• ILSVRC-2012 results table and challenge page, image-net.org. Source for every leaderboard figure, the per-entry method descriptions, the extra-training-data annotation on the 0.15315 entry, the localization results, the 30 September 2012 23:00 GMT deadline, the 12 October 2012 workshop at ECCV in Florence in conjunction with the PASCAL VOC workshop, the 13 October results release, the dataset sizes, and the organizer list.
• ECCV 2012 workshop programme, eccv2012.unifi.it. Confirms ECCV 2012 ran 7–13 October 2012 in Florence and that the PASCAL VOC2012 workshop was held on Friday 12 October. The ILSVRC session is not separately listed there; the Florence date in this entry rests on the ImageNet challenge page's own statement, cross-checked against this programme.
• **Geoffrey E. Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, Ruslan R. Salakhutdinov, Improving neural networks by preventing co-adaptation of feature detectors, arXiv:1207.0580, submitted 3 July 2012.** Source for the dropout date, authorship and abstract.
• **Karen Simonyan, Andrew Zisserman, Very Deep Convolutional Networks for Large-Scale Image Recognition, arXiv:1409.1556, September 2014.** Source for the LRN refutation, quoted from Section 2.1, and for the 16–19-layer depth claim from the abstract.
• **Andrej Karpathy, What I learned from competing against a ConvNet on ImageNet, 2 September 2014.** Source for the 5.1% figure and its construction, the second annotator's ~12%, and the "tradeoff curve" caveat.
• **Olga Russakovsky, Jia Deng, Hao Su et al., ImageNet Large Scale Visual Recognition Challenge, arXiv:1409.0575.** Source for the 1,281,167 training image count and for the statement that "the vast majority of entries in 2013 used deep convolutional neural networks." Section 6.4's human-accuracy numbers were not readable in the version reachable here; the human figures in this entry are taken from Karpathy's own account instead.
• ILSVRC-2013 results, image-net.org. Source for Clarifai at 0.11197 and the ZF entry at 0.13511.
• **Vinod Nair, Geoffrey E. Hinton, Rectified Linear Units Improve Restricted Boltzmann Machines, ICML 2010, Haifa, 21–24 June 2010, pp. 807–814.** Bibliographic detail only; the paper was not read for this entry.
• NVIDIA financial results. FY2013 (ended 27 January 2013) revenue $4.28 billion, up 7.1% from $4.00 billion, from NVIDIA's own annual results announcement. FY2026 revenue $215.9 billion and data-centre revenue $197.3 billion, from NVIDIA's fourth-quarter and fiscal-2026 results announcement, February 2026.
• **Jürgen Schmidhuber, 2011: DanNet triggers deep CNN revolution, people.idsia.ch.** Source for the four DanNet contest dates, the 0.56% versus 1.16% traffic-sign figures, and his characterization of AlexNet's relation to DanNet. Quoted as the claimant's own statement of his own priority claim, which is how it should be read.

Secondary, and marked as such because this entry's colour depends on it:

• **Computer History Museum, CHM Releases AlexNet Source Code, 20 March 2025.** Source for the release date and licence, the five-year negotiation from 2020, Google's ownership via DNNresearch, the bedroom-at-his-parents'- house detail, the origin account (Sutskever persuading Krizhevsky, Hinton as PI, cuda-convnet extended for multiple GPUs), and the Hinton quotation about the Nobel Prize.
• **IEEE Spectrum, How AlexNet Transformed AI and Computer Vision Forever, 2025.** Corroborates the CHM account.
• **Timothy B. Lee, Why the deep learning boom caught almost everyone by surprise, Understanding AI.** Source for Hinton's Nvidia email quotation and for the LeCun "unequivocal turning point / This is proof" quotation as reported.
• **Cade Metz, Genius Makers (2021), via secondary summaries.** Source for the DNNresearch auction narrative and the $44 million figure. The book was not read directly. A first-hand participant account circulating publicly gives $42 million; the discrepancy is recorded, not resolved.
• **Dave Gershgorn, Quartz, The inside story of how AI got good enough to dominate Silicon Valley, 2018.** Source for Krizhevsky's September 2017 departure from Google and his stated reason.
• **Nature, The 25 most-cited papers of the 21st century, 2025**, via University of Toronto's summary. Source for the eighth-place ranking and the 198,000-citation figure as of early 2025.

Not obtained:

• A primary transcript or recording of the 12 October 2012 session, or of LeCun's remark. Google+ is defunct and web.archive.org is not reachable from this job.
• The PASCAL VOC 2012 workshop programme at host.robots.ox.ac.uk, which refused connections on every attempt, and which would have confirmed the session ordering directly.
• Fei-Fei Li, The Worlds I See (2023), the apparent origin of the LeCun quotation.
• The price actually paid for the 3 GB GTX 580 board-partner cards, and how they were funded. The $499 figure in this entry is the reference card's launch MSRP and is not what the training rig cost.