news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

The west wind blows from Aofei Temple
Quantum Bit | Public Account QbitAI

ICML 2024 Test of Time Award announced, Yangqing Jia co-authored the award!

The paper is titled “DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition” was completed by Jia Yangqing and his team while at UC Berkeley 10 years ago.

ICML officials immediately tweeted their congratulations, and Jia Yangqing responded, "I am deeply honored that DeCAF won the ICML2024 Test of Time Award. This has been an amazing decade for the development of artificial intelligence," and tagged all the other authors.

One of the authors, current vice president of Google DeepMind, co-leader of the Gemini projectOriol VinyalsHe also said on the mic, "Thank you very much for this award (it makes me feel old)":

DeCAF is the first open source version of AlexNet, where we tested whether the features learned by this remarkable ImageNet classifier are broadly applicable to other vision tasks. It turns out that this idea is still very relevant to the best multimodal models today!

The two old partners also reminisced about the past in the comment section:

This year is the 41st ICML, with over 9,000 submissions and an acceptance rate of 27.5%. In addition to the Time Test Award,The best paper awards have also been announced, including the Stable Diffusion 3 paper。

This paper, which won the Test of Time Award, has been cited 6,012 times on Google Scholar:

Let’s first take a look at what this paper talks about.

The predecessor of the famous framework Caffe

This paper proposes a method calledDeCAFThe deep convolutional activation feature (Deep Convolutional Activation Feature) is used to solve general visual recognition problems.

The main purpose is to explore whether the intermediate layer features of deep convolutional neural networks pre-trained on large-scale labeled datasets such as ImageNet can be effectively transferred to other visual tasks, that is,Feasibility of transfer learning。

The authors adopted the 2012 proposal by Geoffrey Hinton, Alex Krizhevsky, and Ilya SutskeverAlexNetConvolutional neural network architecture, including 5 convolutional layers and 3 fully connected layers. After pre-training on the ImageNet dataset, the network weights are frozen.

We extracted activations of different layers as features, specifically DeCAF5, DeCAF6, and DeCAF7 (activation of the last hidden layer), and trained only simple linear classifiers on the new task while keeping the DeCAF features unchanged.

To validate the effectiveness of this approach, the authors conducted experiments on several standard computer vision benchmarks, including object recognition (Caltech-101), domain adaptation (Office dataset), fine-grained recognition (Caltech-UCSD Birds dataset), and scene recognition (SUN-397).

Experimental results show that DeCAF achieves excellent performance on all these tasks, often surpassing the best methods at the time.

The t-SNE algorithm is used to map high-dimensional features into 2D space, demonstrating that DeCAF features outperform traditional features such as GIST and LLC in semantic clustering.

Even in the case of scarce training samples, such as single-sample learning, DeCAF still performs well. The paper also analyzes the computational time distribution of each layer of the network in detail and finds that the fully connected layer takes up most of the computational time.

In addition, the paper also explores the impact of regularization techniques such as dropout, especially when applied to DeCAF6 and DeCAF7 layers.

Finally, the authors open-sourced the DeCAF feature extraction tool and pre-trained model.

After seeing this paper win an award ten years later, some netizens suddenly realized, “Is this the origin of Caffe?”

Jia Yangqing also responded:

DeCAFTraining is not fast enough(We estimated that training would take more than a month) so we switched to Caffe. That’s why one has zero caffeine in its name and the other has — caffeine makes things faster for both humans and machines.

Best Paper Award

In addition to the Test of Time Award, the ICML 2024 Best Paper Award has also been announced. This year, there are 10 winning papers.

These include the Stable Diffusion 3 paper "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis”。

Pika co-founder and CTO Chenlin Meng participated in theDiscrete Diffusion Modeling by Estimating the Ratios of the Data DistributionThis work also won an award.

Reference Links:
[1]https://icml.cc/virtual/2024/awards_detail
[2]https://arxiv.org/abs/1310.1531
[3]https://x.com/jiayq/status/1815653822028738667
[4]https://x.com/jiayq/status/1815862939569774796