Thought-provoking article on AI
The New York times (who else? – they really seem to be all over this machine intelligence stuff lately) has published an Op-Ed by Jaron Lanier, a research scientist at Microsoft. The article twists the typical conversation about AI on its head, saying that one of the effects of AI is that people are turning more machine-like rather than machines turning more human-like, and in the process we are essentially losing our soul.
Pretty heavy stuff, for sure.
I’ll focus on one interesting snippet in the article:
We must instead take responsibility for every task undertaken by a machine and double check every conclusion offered by an algorithm, just as we always look both ways when crossing an intersection, even though the light has turned green.
This passage kind of struck a personal chord for me. In my former job, we worked on a system called TCAS, which is a collision avoidance system mandated by law to be installed on all aircraft around the world of a certain size. Certainly, all the commercial passenger jets in the United States have this system installed.
This system could easily be called artificial intelligence – it uses sensors to detect approaching aircraft and issues commands, based on decisions made by algorithms, to the human pilot who would nominally be expected to execute the commands faithfully. The system was developed because of an act of Congress after a series of tragic mid-air collisions in the 50′s, 60′s and 70′s. The system “went live” in 1992, I believe, and there are international bodies containing very dedicated and smart people continuing to work on upgrading it since then.
Where it gets interesting, and related to the author’s quote above, is where the human fits in with this system. For example, pilots are supposed to always follow TCAS’ commands, unless they have overriding visual evidence to support a different decision. The reason is that multiple TCAS systems will actually coordinate actions between the two planes if necessary to separate them.
Now, where it gets really interesting (and tragic) is when despite the rules, one or both pilots choose to take actions opposite to what TCAS tells them to do, as they did in Uberlingen Germany in 2002. In this case, one of the pilots obeyed the commands of the controller, and as a result, there was a mid-air collision and 71 people died.
While there are indeed technical solutions (CP112E) in this case – which I personally worked on verifying – it brings up a case where human judgement was faulty and the algorithm was correct, and the result of the human “double check” caused harm.
In general, I am actually in the author’s camp that handing over your free will to a machine should at least give you pause, and I don’t dispute that we are slowly becoming more machine-like. I support all the creative/independent thinkers out there, because I think it’s getting tougher for people to trust your judgement in the presence of the Social Network.
But I also believe that in some situations, like in Uberlingen, it’s not always a bad thing to accept a machine’s advice. Unquestionably, the human element introduces “noise” which waters down the effectiveness of algorithms such as TCAS. I predict it will be the great ethical and legal struggle of my generation to find out where to draw the line. While it would be great if it were always an individual choice (such as shutting down your Facebook account), I can see examples similar to Uberlingen coming down the line in other life or death situations like in medicine and for safety systems in other contexts where something will have to give.
I seem to be more optimistic than the author, however, on the macro level. I think the slow revolution of society being caused by the AI all around us is causing us to think about our lives and society more logically which will ultimately lead to more social justice.
Either that or you’ll get the computer from Wargames, not sure.
Tic-tac-toe anyone? The only winning move is not to play...
Information/Decision systems in retail
Sam’s club recently started offering their customers personalized deals based on their past buying history:
Linda Vytlacil, vice president for member insights and innovation at Sam’s Club, said coupons normally had a response rate of 1 percent or 2 percent. With eValues, she said, as many as 20 percent to 30 percent of eligible customers collect the discount they are offered.
The program is called “eValues” and it is the “latest iteration in the fast-growing field known as predictive analytics, which uses vast amounts of data to spot trends and anticipate consumer behavior.”
In a related article in the New York Times (yes, occasionally I do read other stuff) there was a description of how people are being offered product deals to share their personal data on sites like mint.com and foursquare.
While taking Intro to Marketing this past semester, I was struck by how much the field could benefit from quantitative techniques like machine learning and decision/control theory, and how the best retailers (like Wal-Mart) have already harnessed the concept of real-time/individualized data analysis. The push only appears to be accelerating. After all, as a shopper, don’t you want the best deals on things you actually want to buy? It seems like a win-win.
I imagine the trick would be to judiciously use the data to “nudge” people to buy things they normally wouldn’t buy – isn’t that the point of sales in the first place? That could get pretty tricky, but I can see how the data could be used to do that in an optimal sort of way. For instance, you could use the data to group similar products and then predict whether this individual is likely to stray from his/her preferred brand on that day and then give the perfect coupon tailored to that person for a rival brand.
Of course, if that fails, free samples always work…
I prefer the chicken teryaki...
Announcement – New LGO EECS track: Information and Decision Systems
The Information and Decision Systems (IDS) track is designed for LGO students in EECS who want to both explore and develop practical skills in how to apply the latest algorithms and mathematical analysis in real operational settings.
The goal of the track is to make LGO the premier training program for leaders who will use advanced data analysis to make smarter operational decisions. The track includes five courses in four areas: 2 courses in tools/theory, 1 course in design, 1 course in communication and 1 course in an engineering elective specific to an application area.
We have a great adviser in Professor Patrick Jaillet, who is associated with LIDS and is also the new co-director of the Operations Research Center at Sloan.
In the coming months, we will be preparing track materials for the LGO Open House and fleshing out how to make this track an active group in the LGO community, much like the sustainability program in ESD. Shoot me an email if you are interested or have ideas…
Google and IBM say we need to train more supercrunchers
There was an article in the New York Times today about the effort that companies like Google and IBM are making to allow university students access to very powerful computing environments to allow engineers and scientists to plow through massive data sets. Their argument is that students are being trained right now to think on a gigabyte scale (if they’re lucky enough to be trained how to analyze real data at all), when all the breakthroughs are happening with datasets in the tera and peta-byte scales.
I couldn’t agree more with this analysis. If people are serious about analyzing those “very rare events”, “long tails” or whatever that can make the difference between a profit and loss, success or failure, or even life or death, then we can’t continue running around assuming things because the model fits 80% of the time and anyways, it’s too hard to do that level of analysis. We all saw what happened with that idea.
When I was working at Lincoln, we created a highly accurate model of U.S. near mid-air collisions. We did this by analyzing about 5 terabytes worth of radar data from across the country (about 8 months worth). Nobody had ever done this before on anything close to that scale.
As a result, we had orders of magnitude more data on near mid-air collisions (a very rare event) than the last model in the early 90′s. Without this data, and the high-powered systems available at Lincoln that we used to analyze it, our model would have suffered from the same assumptions and modeling error as previous attempts, and that is just not good enough for developing something as important as the next generation of collision avoidance systems for manned and unmanned aircraft, which people are now doing at Lincoln, largely as a result of that effort.
The ability to analyze massive data sets has been proven again and again as a competitive advantage in bio-tech, finance (those who do it correctly), internet, and even marketing, making those companies who developed those competencies hundreds of billions of dollars.
Is it then a stretch to say that the next lucrative opportunity in operations management will be to develop the capabilities to harness the massive amounts of data companies already generate every day? I’m talking about everything from inventories to machine control outputs and even to intra-company emails. There are signals in that data, just as there are signals in everything from our DNA to the stock markets, if you look hard enough.
To be honest, I don’t know (I’m new to this stuff!) but that’s why I and several of my classmates are trying to start a new track for LGOs in the EECS department this year called Information and Decision Systems. The focus in this track is to develop the theoretical, practical and communication skills for students who want to take on this operations challenge in the real world, for real companies. That means not just studying and learning the algorithms, but also getting a design background in the networking, database and parallel computing systems that are critical enablers of this type of work. It also means developing specialized communication skills to explain the opportunities and the results, because like the NYT article said, most people have not been trained to think on this scale before.
I could talk for pages more about this topic, but lets just leave it at that for now. I just had to write something because I’m obsessed with this idea, and this article got me all excited. I’m definitely going to look into Hadoop…
Plug for MIT Lincoln Laboratory (“The Lab”)
As cool as it looks
For the past two years I have worked at MIT Lincoln Laboratory, or the Lab for short. Not many people outside of the research circles know about this place, and in fact I only stumbled upon it in my job search because I went to a job fair for high tech companies and got some literature. In fact, the wikipedia page on the Lab is comically short considering the amount of research that has been conducted there for the past 50+ years – while the Lab didn’t invent radar, it probably perfected it. However, because most of this research is classified for national security, 95% of the amazing work by Lab scientists goes unpublished. When I was looking around for information on the Lab, I could hardly find any online, so hopefully somebody may stumble upon this page when they are thinking about working there.
While at the Lab, I worked for the Surveillance Systems group, Group 42. When I got to the Lab, the group was called the Air Traffic Control and Surveillance Systems Group, and about half of the group’s research is sponsored by the FAA. Most of my work centered around the TCAS collision avoidance system, which is mandated by Congress to be installed on every commercial passenger aircraft above a certain size.
TCAS vertical speed indicator (displayed in cockpit)
I also worked on new collision avoidance systems for UAVs (Unmanned Air Vehicles). My major project was developing airspace encounter models for generating random, realistic encounters so that these systems can be tested in simulation. Over the course of developing these models, I learned a great deal about Bayesian networks, Monte Carlo methods, importance sampling, radar (note: the giant radar ALTAIR in that link is operated by Lincoln for missile defense in the Marshall Islands), flight dynamics, and especially air traffic control. It was a fascinating project, and only one of many fascinating projects that I was involved in (I will probably blog about my thoughts on ATC at some point in the future).
Global Hawk, one of the platforms we worked with.
I’ve worked with some amazingly smart and capable people who I now consider my friends. In particular Mykel Kochenderfer and Jim Kuchar have been my mentors, among many others who don’t have personal websites. Over half of the Lab employees have PhDs, most from MIT and other top schools, and the sheer brain power there is kind of awesome. People also work at the Lab for the love of their research, and that definately comes through as well.
I will be leaving the Lab in a couple months because I accepted an offer to be a LGO fellow at MIT. However, my work there has been very intellectually satisfying, and it has served to focus my future academic interests. For instance, I will recieve a MS in Computer Science, largely because my work at the Lab has piqued my interest in AI and machine learning. In particular, I hope to research how these principles can be applied to improve the efficiency and operation of manufacturing companies. I would say that the Lab is an excellent place to work if you are interested in doing cutting edge research, are intellectually curious and smart, and enjoy tackling difficult problems in the area of national defense, homeland security or the FAA.
