How Facebook AI Engineers use AutoML to teach FLOW to make Asimo

Facebook is really serious about AI (Artificial intelligence).

So much so that they’re using AI to build more AI programs as reported in the article “Building AI is Hard - So Facebook is Building AI that builds AI”, published 05.06.16 by Cade Metz, Wired.

The idea originated from Facebook software engineers and basically involves working on an automated machine learning engineer AI called Asimo. This idea of creating an AI to build other AI for advertising purposes is a natural fit to Facebook.

Another way to look at it is that it’s basically an AI that builds other AI systems, like a ship-shipping ship shipping ships.

Already, Facebook uses an AI to parse and analyze satellite data for places where habitation may be located as explained in my blog article entitled “How Facebook's AI is scanning Satellite images to connect 10% of the World without Internet”. 

This comes as no surprise as companies like Google and Facebook can only find a few people who are experts at building AI, which is the equivalent of building a ship. Like shipbuilders, they have to make the machine learning algorithms that will be used in AI and chatbots.

They then then test it by exposing it to real world data as was the case with Microsoft’s AI Tay.ai as explained in my blog article entitled “Why Microsoft Tay.Ai and Facebook chatbots are is future of Fast food and online shopping”, just like building a ship as illustrated in the video below.

These AI shipbuilders are very expensive and cost top dollar to employ. They’re also stretched to the limit running trial-and-error testing on various AI and often are employed by various companies aside from being faculty at certain prestigious universities.  Since this isn't really possible to expose an AI to ever possible situation it will encounter, they usually run simulations in supercomputers to generate the data.

Using various software tools, some of which these AI software engineers build themselves, the the reactions of the AI engine to external stimuli can be judged. This is a very time-consuming activity, as the data dump from the AI's servers can be in petabytes and may often be meaningless metadata.

Facebook goes with the FLOW - Production line for testing AI

Hussein Mehanna and other Facebook engineers on the Facebook ads use bots to improve the selectin engine for ads that appear on your Facebook page. They also have plans to make these AI building tool open source, allowing other developer to get in on the development of AI for Facebook.

Bots are a big deal, as they’ll soon be replacing all those apps on your smartphone as noted in my Geezam blog article entitled “How Facebook Messenger will replace your smartphone in 2016”. 

Bots are currently the focus of Facebook, Google and even Microsoft, who are all betting big on them replacing the pantheon of aps that live on your smartphone.

They're also the ones behind the chatbots that are now invading your Facebook Messenger feeds while you talk to multiple persons as noted in my Geezam blog article entitled “How to use Facebook Messenger to talk to more than 50 people”.  To do this, they employ a lot of AI software tools, such as FLOW.

FLOW is a tool that helps build, test, and execute machine learning algorithms that will be used in AI and chatbots on a massive scale across multiple servers. On the most basic level, it's really a feedback loop that tests machine learning algorithms to see if they can make the final cut to become chatbots and AI that Facebook software engineers can use for advertising and other purposes.

Using FLOW, Facebook can now test some 300,000 machine learning models monthly, increasing their throughput to multiple AI per week instead of one every sixty (60) days.

Some of these machine learning algorithms eventually make the cut and become AI and chatbots. A few of these chatbots end up in Messenger where corporate clients can use them to have conversations with you to sell you everything from insurance to free pizza coupons as noted in my Geezam blog article entitled “How FB Messenger Ads means Chess playing Chatbots with Fast Food Coupons”. 

FLOW automates the testing of the AI, and does this on its own, testing all possible types of data using logistic regression and other testing methodologies. The test data is then represented on decision trees that reflect what the machine learning algorithm being tested is “thinking”.

Some of these AI can then be used for various tasks normally reserved for humans such as:

1.      Add caption to Facebook Videos

2.      Choosing the links for your Faceboook News Feed

3.      Generate audio captions for photos

4.      Recognize faces in photos posted to the social network

So where do you get input stimulus to test the machine learning algorithms that will eventually be used to make AI for chatbots?

Facebook AI that test machine learning algorithms - AutoML teaches FLOW to make Asimo

Facebook engineers have built another AI testing tool called AutoML that analyzes the results from the FLOW tool.

Based on these results, it can then optimize the simulations that FLOW generates to test these new machine learning algorithms. Even more interesting is that this optimized data generated by AutoML from FLOW testing can then be used to train another machine learning model.

This model can then be fed back to FLOW so as to optimize its training of future new machine learning algorithms. This is effectively a feedback loop between an AI trainer, FLOW and the AI data analyzer, AutoML.

In so doing, this makes FLOW function more efficiently and reduces the time to discover algorithms and parameters that are likely to work.  Eventually, they'll make Asimo, an automated builder of AI using that data generated by the feedback loop interaction of FLOW and AutoML.

If Hussein Mehanna and other Facebook engineers plan to make these AI building tool open source are true, then you’ll really need to watch out for your refridgerator coming to life in the middle of the night.

As I said before, a ship-shipping ship shipping ships!