During the opening F8 2018 keynote Facebook CEO Mark Zuckerberg showed the latest updates of the company's Instagram: Spotify integration, AI-based anti-bullying comment filters AR camera effects and four-way video chat.
During Day 2 Keynote, Facebook revealed how its daily Instagram updates are giving their AI technology an intensive course of deep learning in image recognition, one that apparently He made his AI even smarter than Google in categorizing objects in photos.
Facebook achieved this, surprisingly, by instructing its AI to read photo tags and interpret the theme of the photos.
Using this strategy, called "weakly supervised training," Facebook's AI achieved a record high of 85.4% accuracy in an industry-wide image recognition test, surpassing Google's previous record.
A blog post from Facebook Engineering went into detail about the methods. Facebook gave the AI 3.5 billion Instagram images to learn, tagged with 17,000 hashtags for categorization.
Then, Facebook trained AI to put aside non-visual or vague labels and focus on those that offer the most specific categorization.
The team was able to train their AI using this method in just a few weeks by distributing the process on 336 GPUs at a time. Compared to the usual method of teaching AI to categorize photos using visual cues, this method was much faster and less laborious.
A "supervised learning process often produces the best performance results," the post says, "but by hand" the tagged data sets are already approaching their functional limits in terms of size. Climbing up to billions of training images is unfeasible when all supervisions are supplied by hand. "
Thus, while the object recognition software may discover that a subject of a photograph is a bird, the AI of Facebook will use labels or legends to figure that it is, say, Meadowlark.
AI from Facebook
Some important things to keep in mind, in light of the recent Facebook privacy scandal : Facebook engineers only used public photos to train their AI, no private accounts.  In addition, artificial intelligence focuses on object recognition, not facial recognition, so it will not use hashtags to discover which of its Facebook friends is its #bff.
Instead, it will use this information to improve automatic audio subtitles for people with visual disabilities, and engineers also plan to "use artificial intelligence to better understand video sequences or change the way that an image is classified in the sources of Facebook ", or" improve the way in which memories are resurfaced in Facebook ".
In general, Facebook artificial intelligence technology has been a major focus of F8 2018 . On Tuesday, Facebook revealed that its assistant Messenger AI could now automatically translate messages in other languages.
And today, Mike Schroepfer, Facebook's chief technology officer, said in F8 that the company uses AI to automatically remove spam, fake accounts and Facebook propaganda content, including "two million of pieces of ISIS and al -Contenido de Qaeda ".