ImageNet

In 2012, AlexNet won the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) by a margin of nearly 11 percentage points over the second-place entry, reducing the Top-5 error rate from 26% to 15.3%

AlexNet's victory in the ImageNet competition officially ushered in the era of deep learning.

Deep learning achieved a breakthrough performance in the 2012 ImageNet competition, driving the revival of connectionism

The transfer learning strategy uses ImageNet pre-trained weights as a foundation and fine-tunes on agricultural-specific datasets, enabling high recognition accuracy to be achieved with relatively little training data.

In 2012, Hinton's team's AlexNet won the ImageNet competition by an overwhelming margin.

Traditional AI evaluation criteria often focus on model parameter count, inference speed, or benchmark scores.

The choice of pretraining (ImageNet-1K vs ImageNet-21K, supervised vs self-supervised like MAE/DINO) significantly impacts downstream performance, potentially causing several percentage points of difference

The construction of ImageNet relied on nearly 50,000 annotators from 167 countries on the Amazon Mechanical Turk crowdsourcing platform, at a total cost of approximately $22 million.

Mainstream benchmark datasets (such as ImageNet and COCO) consist mainly of natural scenes and modern objects; models perform excellently on these data but struggle to generalize to cultural heritage images with complex textures, uneven lighting, and scarce annotations.

On vision tasks, UTA also improves the performance of ResNet50 and ViT-B on ImageNet.

The SPOOK experiments evaluated OpenCLIP and DINOv3 models on the ImageNet and Flickr30k datasets.

Open-source frameworks PyTorch and MONAI as well as pretrained weights ImageNet and RadImageNet can be used to save training time in a hackathon.

Inception V1 (GoogLeNet) was the winning network of the 2014 ImageNet competition, introduced the Inception module and 1×1 convolution dimensionality reduction, and is about 22 layers deep.

Inception V1 is a classic convolutional network proposed by Google in 2014 and, due to its breakthrough on ImageNet and its compact structure, has long been the "fruit fly" model for interpretability research; its conv5-1 layer has 480 channels and was systematically studied by Olah et al. using feature visualization.

Training only the classification head (freezing the entire backbone) suits scenarios with a sample size in the hundreds to thousands and with classes semantically close to ImageNet, but the model's ability to adapt to domain shift is almost zero.

MFlow is evaluated on ImageNet, UCF-101 video, and large-scale text-to-image generation tasks, and the results show competitive or superior generation performance across a variety of settings.

Through empirical studies on different architectures (such as fully connected networks, convolutional networks, and residual networks) and datasets (such as MNIST, CIFAR-10, and ImageNet subsets), the authors found that the training process continuously expands ray coverage, especially in directions where the Jacobian singular values are smaller.

Guo et al. in 2017 found that deep models on ImageNet had average confidence far higher than actual accuracy, and temperature scaling is a simple and effective post-processing calibration correction method.

MiniDog's pre-training uses 26,000 dog images from ImageNet, and generates text descriptions for them.

The paper tested GCD on CIFAR-10 and ImageNet, covering three architectures, ResNet-18, ResNet-50, and ViT-B/16, and evaluated four mainstream attack methods.