For a few day two, is really what F8 is about. People tend to forget that F8 is a developer conference. Day one was structured heavily on product feature announcements which appealed to the larger press and people.
However, day two is what most developers look forward to. It’s in-depth technical talks on what Facebook is doing, how it’s doing it and what we can learn. Day two had a keynote, live classrooms, panels and in-depth sessions on some of the behind the scenes research that Facebook is conducting.
I have attempted to recreate some insights bottom-up. A heads-up, the nature of the sessions, demands a bit of technicality to be explained, so please bear with me. If you prefer to just see the highlights, I suggest you read the name of the bullets in bold, followed by the 'why is this important' section. There are three key pillars that one could reconstruct.
1) AI for online safety
2) AI for inclusiveness
3) AI for commerce
It is important to note that one AI technology can help all of the above, so it’s not separate endeavors, rather different applications of the underlying technology.
Let’s look at the technologies now and how they impact each of these areas:
1) Computer vision: Computer vision is a broad category of AI technologies that allow machines to understand images and video. Machines only understand pixels and hence, it’s harder than it appears to train them to understand images. Facebook has done some stellar research in this category in the past four years with their FAIR (Facebook AI Research) team. The current best image/video benchmark is set to be at 77.7 per cent accuracy. Facebook introduced a new AI model that has increased that world benchmark to 82.8 per cent. That’s a significant improvement in the AI world. It translates to a 25 per cent lower error rate. They achieved this by training on 65 million public Instagram videos and their related posts. The previous model had trained on just 6 million in comparison.
Why is this important?
With better image/video recognition, Facebook combats hate speech, pornographic content and general spam. In just Q3 of 2018, Facebook’s AI removed 1 billion spam messages and 720 million fake accounts. Understanding image content, helps Facebook have more robust systems in place than by just relying on text. From a commerce point of view, the better Facebook’s AI understands product images say in Marketplace, the better suggestions (read sponsored ads) it can be surface to you, hence increasing your chances to buying something.
2) Machine translation: There is a huge challenge in the creating machine learning models for translations, because it requires you to have a lot of 'paired data'. Basically, you need a large file, with one line in English and say one line in Spanish that mean the same thing, paired one after the other. Of course, such data is not naturally occurring, and companies commission such data generation activities or buy data sets from other research institutes that have created such sets. Another issue is that for lesser spoken languages, there is almost no such data and could take years and millions of dollars to build. Facebook has managed to overcome this by a rather innovative method. I won’t be able to get into the details of it due to the depth and technical drill-down that it requires, but suffice it to say, that its rather ingenious and using existing public post data to create a 'universal translator', that can translate between any two languages directly, without needing to convert it to say English first.
Why is this important?
There are over 6,000 languages being spoken on earth. Facebook’s mission is to the connect the world. When people speak different languages, it creates a barrier to connecting with them. Translation is the key to inclusiveness from a Facebook perspective. It also improves safety on the platform, by being able to understand more languages and put checks and balances in place more proactively. It’s also currently being used for the Indian Elections supporting 22 languages to keep fake news in check and also take down hate speech.
3) Conversational AI: There were a number of interesting talks around conversation and the research that Facebook is powering to make conversations with machines and bots more natural, intuitive and powerful. The challenge Facebook is trying to solve is how to move away from the single task interaction with have with assistants. So, for example, today you would have to say, 'Switch off the lights' and 'Turn on the music' in two different commands, rather than just say 'turn on some music while dimming the lights'. The latter confuses the assistant, since it contains two tasks. There is some early research that Facebook shared on how it is improving its understanding of being able to build bots that can carry this out in the future.
Why is this important?
While the Amazons and Googles of the world feel voice is the future of the user interface, Facebook has taken the messaging route instead. It makes sense, teens text more than they speak. Today already smart marketers are using messaging to acquire leads, nurture them, get them to sign up and even keep them engaged post purchase. With a combination of bots and a new age contact center, the dynamics of customer interaction have changed drastically. I see Facebook going quite big on conversational commerce as their strength and unique positioning the market.
Of course, these three are not the only things Facebook spoke about. There was a lot of buzz of SparkAR and how it could change the current Augmented Reality possibilities. Then there is Portal and Oculus on the hardware side. Overall, its been a super eventful F8 and its very exciting to see where technology at large is headed with torch bearers like Facebook.
(The author is CEO, DAN Programmatic and chief data officer, DAN South Asia.)