Do you know Facebook is ahead of the race to build AI deep learning models because of a secret weapon that other’s lack; Your billions of public Instagram images. This announcement has been made on Tuesday 1st May, at Facebook’s annual F8 developer conference.
Facebook is using each photo that you are sharing on Instagram either it is a plate of Pizza or a breathtaking location, to train its artificial intelligence algorithms for understanding objects in images. The hashtags that are provided with the billions of images are a major source for training the AI models as the images are already annotated and no manual effort is required for analyzing and annotating the data. The trained model that is available at the end is beating the industry benchmarks in this field.
According to Mike Schroepfer Facebook’s Chief Technology Officer, “We rely almost entirely on hand-curated, human labeled datasets. If a person hasn’t spent the time to label something specific in an image, even the most advanced computer vision systems won’t be able to identify it”. But the availability of labeled Instagram images made the task of collecting and training object recognition and computer vision models relatively easier and provided an edge over the other big competitors like Google and Amazon. It is evident from the results achieved by facebook. Schroepfer further added, “We have produced state-of-art results that are 1 to 2 percent better than any other system on the ImageNet benchmark“.
About 17000 hashtags have been extracted from 3.5 billion public Instagram posts and these images have been used to train AI models that will eventually be used to categorize images with a higher accuracy without any human intervention. According to the CEO Mark Zuckerberg, AI is going to help Facebook in removing extremist or violent content that has been shared by the users. Facebook is already working to eliminate nudity and violence from the videos and images shared its users.
This approach has been used in the past by other companies but it also raises major privacy concerns. Billions of users are sharing their valuable information i.e text and images on a variety of apps like messenger, WhatsAp, and Instagram, that is working on one platform Facebook. That valuable information collected from these apps can be used to train AI models and eventually it might be posted publicly. Most of the users will not be aware of the fact that the data shared by them on a public platform is being used to mine AI systems.
At the moment, Facebook can only extract object-based data using the AI models. Facebook does not intend to use your photos to draw inferences about your behavior. Currently, about 20,000 human moderators are working to remove bad content and actors from the Facebook. The main aim of using AI-based systems is to improve its moderation capabilities on a huge scale.
“Until very recently, we often had to rely on reactive reports. We had to wait for something bad to be spotted by someone and do something about it. Now the bulk of moderation is being handled by AI, which is helping the company screen for and scrub its platform of terrorist propaganda, nudity, violence spam and hate speech. This is why we are so focused on AI research. We require new breakthroughs and technologies to solve problems all of us want to solve.” Stated Mike Schroepfer.
AI also requires high computing power along with the availability of large-scale data. It took about 3 weeks to complete the training process, valuable resources that were used during the process include 336 high powered graphics processors and 42 servers that were involved to train 3.5 billion Instagram images. In order to maintain its competitive advantage, Facebook also plans to develop its own machine learning chips that will help the company to beat its competitors in the field.
Source: The Verge