Facebook announces image recognition model 'SEER' learned using photos posted on more than 1 billion Instagram

On March 2, 2021, Facebook announced the image recognition model 'SEER ' that was learned using more than 1 billion images posted on Instagram. At the same time , the library 'VISSL ' that enables 'self-supervised learning' used in the development of SEER is also released.

[2103.01988] Self-supervised Pretraining of Visual Features in the Wild


VISSL · A library for state-of-the-art self-supervised learning


SEER: The start of a more powerful, flexible, and accessible era for computer vision

Many existing AI-based image recognition models employ a technique called 'supervised learning ' that requires learning with manually labeled datasets. However, SEER announced by Facebook this time adopts 'self-supervised learning' that does not require manually labeled datasets, and by using more than 1 billion images posted on Instagram, He said he succeeded in obtaining high image recognition accuracy.

Compare the 'classification system score' of the image recognition ability measurement test using 'ImageNet' with existing image recognition models such as SEER and 'SwAV ', 'SimCLRv2 ' and ' Vision Transformer (ViT) ' announced by Google in 2020. The result is like this. The horizontal axis of the graph shows the number of parameters, and the vertical axis shows the image recognition accuracy (the higher the value, the higher the recognition accuracy), and it can be confirmed that the image recognition accuracy of SEER exceeds the existing image recognition model.

'SEER has shown that self-supervised learning is useful in real-world tasks. SEER is a way to build

flexible, accurate and adaptive computer vision models,' said Facebook.

In addition, Facebook has released the source code of the library 'VISSL' used for the development of SEER. 'We hope that sharing VISSL with the AI research community will lead to an open exchange of ideas and accelerate research on self-supervised learning,' he said. I am.

In addition, since more than 1 billion images posted on Instagram are used for the development of SEER, the American news media CNBC said, 'Instagram users are surprised to know that their images are used for AI research. Probably. ' In response to this anxiety, Priya Goyal, a software engineer at Facebook's AI Institute, said, 'We will publish the details of our research and open it to allow other researchers to train with self-teaching learning. I'm creating a source library, but I don't share images posted on Instagram, 'he says , claiming that Instagram users' privacy is protected.

in Software,   Science, Posted by log1o_hf