Research which is done by the company and presenting the same at the annual F8 developer conference about its latest artificial intelligence project. Facebook is taking the help of the Instagram photos which are uploaded by the users with hashtags, and they use this data to train the image recognition models. The company is now depending on hundreds of GPUs which runs round the clock to collect the data. The best accuracy that has achieved till now is about 85.4 percent on ImageNet.
Mike Schroepfer who is the Facebook chief technology officer, says that the challenges that come to improve the accuracy of the computers to understand certain things about the photo. The biggest problem that they are facing now is about not having enough labelled photos which will help to train the computers to understand the nature of them.
The research which was done on Pre-training is focused on a developing system to find the relevant hashtags. This means that the discovering of the hashtags to learn to give priority over some specific hashtags over the general one.
Schroepfer said Facebook had taken the help of about 3.5 billion Instagram photos along with people hashtags to describe them. The company was then able to “Produce state of the art results” on the ImageNet computer-vision benchmark. Researchers of AI used this as a comparison of their project’s effectiveness concerning another project.
An interview to the Fortune, Manohar Paluri, who is the head of applied computer vision said, the main challenge that they face due to many noisy hashtags on Instagram photos and it put a hindrance to train the company’s computers to learn AI. This means in simple words that most of the people are using a wrong hashtag, for example, a different breed of dog is described as husky in the hashtag.
As per a research paper on this project, the researchers said that they are using WordNet, in which the company can group all the common hashtags with each other to cut down the noise in the photos. This was trained in such a way that Facebook can use it for better image results and accessibility tools. This process has lead Facebook’s computers to distinguish between birds and weather conditions in the photos successfully.