What is machine learning and AI really?
As a software engineer, or just a very lazy person in general, to me, machine learning or more generally AI is programs that can program itself with data — to me at least, what’s most intriguing to me is the ability to build system that makes prediction about human without knowing exactly how human makes decisions.
Admit it, we don’t just make rational decisions. Before I join the force of machine learning, I studied political science and economics and worked for market research consultancy. Both focus very much on theorize/summarize human behavior with data and statistics.
Human behavior is very intriguing indeed, it’s amazing to see how different schools of studies rationalize it. One of my political philosophy class in college, the professor discussed different assumptions people made when they are voting (veil of ignorance etc.). Microeconomics in general or game theory is all about theorizing human behavior with self-interest and delicate math.
I must ask, has anyone been able to draw indifference curve in reality? When I first saw correlated equilibrium, it was the most cool piece of proof I’ve seen about traffic light. Don’t think that applies to New Yorker though, I do jaywalking all the time thinking I am minimizing the deadweight loss, it’s for the common good! Let’s be honest, do you think more about how you can cast a more unbiased vote or that episode you just saw on SNL?
You can probably tell I subscribe to no philosophical or mathematical explanation of human behavior, quite the contrary — I do, just that I am more of a frequentist than bayesian theorist, you’ll see. Have you ever done a market research survey asking you questions like “On a scale of 1 to 10, how would rate movie A?” My reaction was always how would I possibly know? It’s just as random as asking me to pick a number with little to no context. With guilt, I’ve written countless market research reports using survey data that makes no sense to me. Now I’ve probably made it very clear, there’s no way you can possibly rationalize human behavior, even with data.
Hopefully I have fully convinced you how I don’t like any of the traditional approach to rationalize human behavior. There is a new approach though, real data, data generated by us. To determine how you really like a movie, instead of surveys, Netflix has the usage log, when did you watch what movies, did you watch the full movie, did you pause, did you go look for similar movies afterwards? Did you watch a lot of movies alike? They won’t be able to tell exactly why do you like certain movies, but they certainly are able to find the movies you like, to me, that’s machine learning.
Traditionally, if we were to program a movie recommendation system, developers would interrogate the guy from movie department asking exactly the logics to program the system. Usually the system ended up with lots of assumptions (subject matter expert’s false beliefs or not) and lots of if clauses like if person A likes movie A and B then person A must also like movie in category C. Because of the vast amount of data that’s available to us and it’s so easy to spin up 200 nodes EMR cluster to repeatedly interrogate the data instead of that poor guy from movie department, we are able to make personalized prediction on what you might or might not like. What’s wonderful to the developer especially, no more if clauses, we are now designing the interrogation algorithms which is usually more generally applicable, simplicity is usually consider a good trait of software.
Common Myths
- Nobody’s nobody’s big brother
No data engineer or machine learning engineer would possibly care about an individual record (unless it’s an outliers or it throws an error when running). No human even the team who developed the algorithm knows what you like unless they plug your data into the algorithm and generate predictions. - Nope, machine learning or AI programs do not have personality
Some commented that the AlphaGo won the game because he can keep his calm throughout the game. Ask any developer if their program is more calm or agitated, perhaps I’ve never established such intimate understanding with the code I’ve written, even for those who do, I don’t believe they will ever put “calm” on project spec. In general, we don’t build software to be calm, we build them to do things we don’t have time to do. - Nope, the world is not going to be taken over by evil AI
AGI (Artificial General Intelligence) is hard, and even the most renowned researcher in the field agrees. It’s weird when people quote Donald Trump’s world on public health policy, and it’s equally weird to quote Elon Musk on AI is going to take over the world instead of someone who’s building the AI system. To be fair, pro-guns often say it’s not gun’s fault, it’s the people who pull the trigger, hopefully they won’t change their stance just because they’ve seen guns but not the actual code in AI programs.
The purpose of this post is to clarify these myths and hopefully give people closer to my background (liberal art/social science) more understanding about machine learning and AI system in general. Personally I feel very lucky to be working closely with the technology I feel most passionate about. If you are from old-school statistics and quantitative research background, they share lots of similarities, the difference is that now we have literally exabytes data generated daily, and thanks to aws, an ordinary engineer like me can easily spin up hundreds of machines to train algorithms of my likes. Even just a few years back when I was in college, this was unforeseeable. The field is growing everyday, the syllabus developed just a couple months ago is probably already outdated due to new technology available.
Adoption of new technology can take time (see Second Machine Age), and we need people to share the excitement (not the fear, just like everything else) of a great technology growing everyday, have the imagination as I do for the future.