AlexNet 赢得 ImageNet 竞赛

15.3%，26.2%。

ILSVRC 2012 的 top-5 测试错误率把两个数字摆在同一行：AlexNet 的集成模型是前者，第二名是后者。十点九个百分点的差距，大到很难再解释为调参运气。

计算机视觉研究者已经见过卷积神经网络——LeNet 一类网络在更早的数字与小规模识别任务上就存在——也知道 GPU 擅长并行计算；ImageNet 也已提供超过百万张带标签训练图像。条件散落在论文、硬件手册与数据集网站上。2012 年秋天这张表突然说明：当数据、并行计算与足够深的卷积被当成同一个系统来训，得到的不是小幅改良，而是别人用旧配方难以追上的分差。

Krizhevsky、Sutskever 与 Hinton 在 NeurIPS 2012 论文《ImageNet Classification with Deep Convolutional Neural Networks》里给出的网络有五个卷积层和三个全连接层，约六千万个参数、六十五万个神经元。训练图像约 120 万张，另有 5 万验证、15 万测试；图像先缩放到 256×256，再截取 224×224 的输入。分类目标是 1000 个 ImageNet 类别。与竞赛中常见流水线不同，系统不再先手工设计 SIFT、HOG 之类局部描述，再接词袋与支持向量机，而是让卷积核直接从像素统计里形成层级特征。作者并没有写「特征工程已死」；他们把设计重心从「提取什么边缘或纹理」挪到结构、损失、正则、数据增强与硬件约束上。

没有单一技巧能独占功劳。ReLU 使深层网络比 tanh 等饱和激活训练得更快——论文用 CIFAR-10 对照显示，达到相同训练误差所需的迭代次数显著更少。数据增强包括随机裁剪、水平翻转，以及在 RGB 通道上做 PCA 扰动。全连接层使用概率 0.5 的 Dropout。测试阶段对四角与中心共五处裁剪及其水平翻转、合计十个视图取平均预测。硬件限制写进了架构图本身：单块 GTX 580 只有 3GB 显存，完整模型放不进去，于是网络按通道切开，分到两块 GPU 上，只在特定层交换信息。约九十个 epoch，在两块 GTX 580 上训练五到六天。竞赛提交进一步使用七个模型集成，才得到 15.3% 的 top-5 错误率；单模型已经明显低于第二名。论文末尾写得很直白：网络主要受 GPU 内存和研究者愿意等待多久的限制。

竞赛之后数年，视觉研究与工业系统迅速把「手工特征加浅分类器」换成可端到端训练的卷积栈。ImageNet 提供标注规模，CUDA 与消费级 GPU 提供可负担的矩阵吞吐，ReLU、Dropout 与增强提供让大模型可训的细节。ILSVRC 使用 top-5 错误率：五个猜测中只要包含正确标签即算对——这一指标让 15.3% 与 26.2% 的对比集中在同一套规则下。作者公开了架构与训练细节，使其他实验室可以在相近硬件上尝试复现。

AlexNet 赢下的是一次分类竞赛，交出来的却是一组可复做的条件清单：足够多的标注图像、装得下模型的并行硬件、端到端卷积，以及一整套让六千万参数不至于完全过拟合的细节。表格本身不参与争论。它只是把数据、并行计算与深层卷积第一次以无法用修辞缩小的分差，钉在了同一行里。

15.3 percent. 26.2 percent.

In the ILSVRC 2012 results, the first number was the top-five test error of an AlexNet ensemble. The second belonged to the runner-up. A difference of 10.9 percentage points was too large to dismiss as ordinary tuning luck.

Convolutional networks were not new—networks in the LeNet lineage had already solved smaller recognition tasks—and researchers already knew that GPUs could perform parallel numerical work. ImageNet had already made a million-image training set available. Conditions sat scattered across papers, hardware manuals, and dataset sites. In autumn 2012 the table showed what happened when data, parallel compute, and a deep enough convolution were trained as one system: not a modest improvement, but a gap hard to close with older recipes.

Krizhevsky, Sutskever, and Hinton described, in their NeurIPS 2012 paper ImageNet Classification with Deep Convolutional Neural Networks, a network with five convolutional layers and three fully connected layers—about sixty million parameters and 650,000 neurons. Training used roughly 1.2 million labeled images, with 50,000 for validation and 150,000 for test. Images were resized to 256×256, then cropped to 224×224 inputs across 1,000 ImageNet classes. Many leading pipelines still separated the work: local descriptors such as SIFT or HOG, then encoding and a classifier. In AlexNet, design moved from prescribing which edges or textures ought to count toward architecture, loss, regularization, augmentation, and hardware limits.

No isolated trick carried the result. ReLU units allowed optimization to proceed much faster than saturated activations; a CIFAR-10 control in the paper shows the ReLU network reaching the same training error in far fewer iterations. Crops, horizontal flips, and PCA-based RGB perturbations enlarged the effective data. Dropout with probability 0.5 randomly suppressed units in the fully connected layers during training. At test time the model averaged predictions over ten views—five crops and their horizontal flips. The complete model did not fit into the 3 GB memory of one GTX 580, so it was split across two cards, with channels divided and communication only on selected layers. About ninety epochs took five or six days on those two GTX 580s. The competition entry that produced 15.3% top-five error used an ensemble of seven models; even a single model already sat well below the runner-up. The authors wrote candidly that the network was limited mainly by available GPU memory and by how long they were willing to wait.

In the years that followed, vision contests and industrial systems rapidly replaced hand-engineered features plus shallow classifiers with end-to-end convolutional stacks. ImageNet supplied scale; CUDA and consumer GPUs supplied affordable matrix throughput; ReLU, dropout, and augmentation kept a large model trainable. ILSVRC’s top-five rule—count a prediction correct if the true label appears among five guesses—put 15.3 and 26.2 on one shared ruler. The authors published architecture and training detail so other labs could attempt reproduction on similar hardware.

AlexNet won a classification contest and handed back a reproducible checklist: enough labeled images, parallel hardware that could hold the model, end-to-end convolution, and a set of details that kept sixty million parameters from wholly overfitting. The table does not argue. It only pins data, parallel compute, and deep convolution onto one line with a gap rhetoric cannot shrink.