Inside AI - June 9th, 2019 |

Inside AI (Jun 9th, 2019)

Inside AI Weekend Commentary: Goodhart's Law and AI Data Sets

Subscribe to Inside AI

New blank template
Subscribe | View in browser

Happy Sunday and welcome to the weekend edition of InsideAI!  I'm Rob May, CEO of Talla.  I wrote a post this week that may interest you if you like Human-In-The-Loop AI, about our "customer in the loop" model of support automation at Talla and how it leads to 90%+ accuracy for automating customer support inquiries. 

I also host the AI at Work podcast.  The latest edition features Brandon Rohrer discussing data science at Facebook and iRobot.

If you are an investor and want to join the syndicate that invests in deals that come in from readers of this newsletter, apply here.

Here are the top articles from the daily newsletter this week:

Twitter has acquired London-based deep-learning start-up Fabula AI. The company caught the attention of Twitter by developing a simplified model of identifying patterns of disinformation and how it spreads online. Twitter will use Fabula to build out its machine learning capabilities, which will feed an internal research group. That research group will then focus on key strategic areas like natural language processing, reinforcement learning, machine learning ethics and graph deep learning. With increasing political and public pressure to understand fake versus real information, Fabula co-founder Michael Bronstein said he is looking forward to using graph deep learning techniques to improve the health of the conversation. Financial terms were not disclosed.  - TECHCRUNCH

Speaking at this week's Amazon Machine Learning, Automation, Robotics, and Space (re:MARS) conference, actor Robert Downey Jr. announced plans for a coalition that he says will use AI and robotics to "clean up the planet." The "Iron Man" actor — who was one of the inaugural event's keynote speakers — said he plans to launch the Footprint Coalition by April of 2020. While details of the initiative remain scarce, Downey Jr. said the idea is to use AI, robotics and other technology to reduce the world's carbon footprint within the next decade, telling attendees, "[I]f we’ve made even a little dent in what I think is a massive threat to our future and the mess we leave behind, I’m going to come back and throw the nuttiest retirement party you’ve ever seen.” - BUSINESS INSIDER.

Amazon plans to test out AI-driven drones that can deliver household goods within the next several months. Jeff Wilke, chief of Amazon’s global consumer business, made the announcement during this week's Amazon inaugural re:MARS conference. The drones, which resemble a helicopter and take off vertically, are equipped with AI-based sensors that can keep the drone safe and away from other aircraft and people. Other AI-related feats that were on display included an autonomous mower and a Temi robot that followed people around playing pop music. - SEATTLE TIMES

MIT researchers developed a neural network model that can predict what a person's face looks like based on a short audio clip of their voice. In a non-peer-reviewed research paper published on Arxiv, the researchers explain how they used a dataset comprised of millions of YouTube clips to train their Speech2Face model to associate vocal attributes with certain facial features. The AI has the potential for a number of useful applications, such as creating a representative face for people during phone calls, they noted. - FAST COMPANY

-- Commentary --

Two years ago I was at an event in Boston and I happened to sit at a dinner table across from a guy widely recognized to be one of the most brilliant people in the world.  We talked mostly about AI but at one point the conversation turned to hiring, and he told me that for 2 decades he has tracked the performance of everyone who worked for him.  Based on that performance tracking, he had stopped hiring from Harvard and Stanford.  He said that historically, his best employees came from Harvard, Stanford, and MIT, but that starting 5 or 6 years prior, the people coming out of Harvard and Stanford started to really slip in their performance.  I asked why, and he said that he didn't know, but had a hypothesis.  He said "I believe that ivy league college admissions has become so competitive that it rewards people who are good at the admissions process, not people who are good."

This is a form of Goodhart's Law, which says "when a measure becomes a target, it ceases to be a good measure."  You can see where this is going, and that there may be a similar law for machine learning.  We are in a phase of AI where we are using data sets that were created for other purposes, not with AI in mind.  What happens when you know all your data is going to be fed into an AI?  Does it change the data you create?

We see early versions of this now.  You can buy services that help write a resume more likely to get past a resume scanning algorithm, and people use instagram filters to put their best self forward on dating apps - the way they want to be perceived, more than the way they really are.  You can imagine as algorithms run more of our lives, you may buy a service that logs in as you and clicks around online for you to make you seem, to the algorithms, like a certain type of person.  Maybe this data allows you get a better credit score, or find a better quality dating partner, or get a discount on certain products or services, or find a better job.  

Think about fake news, and detecting fake news, and the arms race it can create.  Now extrapolate that to everything about you - particularly those areas where you want to be perceived as something other than your true self.  Data creation may no longer be an innocent thing in a world where every time you do something you wonder "how will this affect the algorithms?"  

Many people have written about the importance of controlling an AGI, should we be able to build it, and the power that would convey.  But, controlling the inputs to that AGI also has power.  Possibly more.  That is how the AGI could be persuaded to make certain decisions.  Or maybe it is harder to fool than the algorithms of today.  I don't know but, it's something I have been thinking about.  If you have ideas, please reply.  I'd love to hear them.  I can't respond to them all but always read every reply.

Thanks for reading.


Copyright ©, All rights reserved.

Our mailing address is:
767 Bryant St. #203
San Francisco, CA 94107

Did someone forward this email to you? Head over to to get your very own free subscription!

You received this email because you subscribed to Inside AI. Click here to unsubscribe from Inside AI list or manage your subscriptions.

Subscribe to Inside AI