One the of challenges with machine learning, according to Whetlab CEO Ryan Adams, is that "it often feels like there's a lot of black magic involved"—lots of levers and slider and buttons and knobs that take an army of PhDs to configure.
Companies like Google,Facebook, and Twitter, have the time and money to spend on tuning such a complex system, but there’s a problem: there's just not enough engineering talent to go around. Now, Twitter is attempting to supplement some of those machine learning engineers with actual machines.
The theory is that, instead of an engineer spending weeks or months getting a spam detection system or a trending topics algorithm working as best as she possibly can, an algorithm could do most of that heavy lifting instead.
In June, Twitter acquired Adams' young machine learning startup Whetlab to work in its nascent Twitter Cortex lab, a hub for AI and machine learning research. The company was founded by five machine learning researchers in March 2014. Whetlab's technology is, essentially, AI for AI (meta, right?).
Think of it this way: With a typical machine learning system, there are lots of knobs and levers and options for engineers to set, and it takes a while to set all of these options in a way that gets you optimal or best results. Whetlab figured, why not free engineers from the tedious, time-consuming part of that scenario and have an algorithm tune and gather insights about all of those settings instead?
"Every time an engineer sets some knobs in a big system, then we provide a big opportunity to set those knobs rationally and to get the very best performance, whether we're talking about a user interface, or energy usage in a data center," Adams told me.
If this sounds pretty abstract, there's potentially a big payoff for users, too: It means that when Twitter mucks about with the systems that handle things such as spam, abuse, or NSFW content—or even the ads and in-case-you-missed-it tweets that get dropped into your feed—there's another watchful AI that's pulling some of the levers and adapting the experience so that it works its best.
The startup is part of a wider effort at Twitter, and other tech companies, to up their machine learning research efforts in recent years. Google, for example, hired renowned researcher and University of Toronto professor Geoffery Hinton in 2013 (Hinton also advised three of Whetlab’s co-founders, including Adams, all alumni of the University). Facebook also created its own AI lab in late 2013, hiring the well-known computer scientist Yann LeCun.
Although it's still early days for the Whetlab team—Adam’s was unable to offer much in the way of specifics on his team’s work at Twitter thus far—there have been some surprises. "Our objective was to automate the process, and to make it less dependant on humans tuning a bunch of knobs in the dark, but it also turned out that humans were not as good at most of the time as AI," Adams explained.
"If you have more than four knobs to turn or so, then humans start to get really bad at this. And it's simply because humans have a really hard time reasoning in more than about three dimensions."
Adams was also pretty clear during our chat that Whetlab isn't replacing Twitter's existing methods of handling spam, NSWF content, or abusive tweets, teams that Whetlab has worked with so far. Rather, Whetlab's technology is meant to augment and improve the way these teams do their jobs. And in the future, Adams believes that there's an opportunity to assist in efforts that even go beyond machine learning systems; pretty much anything that has more than a few knobs and sliders for a human to set is fair game to be optimized.
And if the end result is a nicer, less hostile Twitter? Do your thing, machines. Do your thing.