Facebook succeeded in improving the performance of image recognition by utilizing "hash tags that users attached to images"

Facebook developer conference "F8"It was reported that the performance of image recognition improved by utilizing the hash tag attached to the image by the user like Instagram.

Advancing state-of-the-art image recognition with deep learning on hashtags | Engineering Blog | Facebook Code | Facebook
https://code.facebook.com/posts/1700437286678763/


Developer Conference "F8" 2018: Day 2 Presentation Summary | Facebook News Room
https://ja.newsroom.fb.com/news/2018/05/f8-2018-day2/


Facebook is using billions of instagram images to train artificial intelligence algorithms - The Verge
https://www.theverge.com/2018/5/2/17311808/facebook-instagram-ai-training-hashtag-images

"Image recognition", one of the pillars of AI research, increases accuracy by learning a lot of individually labeled data, but Facebook uses a data set containing 50 million images in the case of Facebook If you increase this to billions of units, it is difficult to tag with manpower any longer.

So, what Facebook attracted as a new breakthrough is a hash tag that the user has independently attached, such as seen in Instagram etc.

byJon Tyson

As a result of training with 1 billion data sets, the accuracy of image recognition in the benchmark tool · ImageNet improved to 85.4%.

For example, it is expected that you can further detail the audio caption for the visually impaired, and quickly discover inappropriate images.