Tuesday, January 19, 2010

Machine Learning in One Sentence

Posted by Danny Tarlow
We've all had this experience: you're out at some party or social event, doing your best to blend in and have a good time. Inevitably, though, somebody wants to talk to you. One thing leads to another, and it happens: they ask, "so... what do you do?"

My default answer is
  • "I'm a Ph.D. student in Computer Science." (optional addition: "I do something called 'machine learning' ")
It's not the most exciting answer in the world, but it's surprisingly effective--from here, it's fairly easy to read whether the person opposite me just perked up at the idea of finding a like-soul or whether their eyes just glazed over as they stumble around looking for the nearest exit. But seriously, it gives somebody the opportunity to press further, but usually it provides a nice opening to change the subject tactfully--it's an improvement over the more extreme tactics that may leave somebody thinking I have some deep dark secret or that I'm a criminal, etc.

This leads me to wonder, though. What, then, is the most exciting answer in the world for a PhD student in machine learning to give? A few rules:
  1. One sentence limit.
  2. It has to be more-or-less true.
  3. (this is the hard one) It needs to be a conversation starter rather than a conversation ender. So while "you don't want to know" satisfies (2), it is disallowed by (3).
I can think of a few not-terribly-creative possibilities:
  1. "I teach computers how to think."
  2. (serious face, matter-of-fact tone) "I develop algorithms for MAP inference in graphical models, typically looking at classes of energy functions where standard techniques like quadratic pseudo boolean optimization or max-product belief propagation are inefficient or don't work well." (Update: this one is meant to be a joke, for the record)
  3. "I pretend my computer is a baby, and I try to teach it about the world."
  4. "I design parts of robot brains."
Help me out here. Do you have better ideas? No need to be a PhD to contribute. As a reward, I promise to take the best suggestion, try it out "in the wild", and report back.

26 comments:

Mark said...

I write programs that make people think computers think.

François Bulens said...

I'm building the early stages of skynet.
With government funding of course.

oakmad said...

You ever wondered how Google knows you mistyped your search?

Colin said...
This comment has been removed by the author.
ephes said...

My work consists of replacing you with a very small shellscript horribly gone wrong...

Stpk4 said...

working to find a way to make computers think like living things

djcjr said...

I try to make computer programs that act a brain.

Winterson said...

"I work on methods for automatically turning data into computer programs, with applications such as web search, spam filtering, and movie recommendations."

And if there's interest, I explain how a spam filter works as the simplest real-world example of machine learning. (You have labeled examples, want a simple function, take words as features, assign some weights, do it all automatically.) Everyone can relate to spam.

But it sounds like your work is in statistical inference, in which case I'd say:

"I take statistical models that describe complex behavior, such as how proteins fold or what products people buy, and try to get the best possible answers out of them as quickly as possible."

The key is to think of concrete, real-world applications and go from there. Be glad you're not in pure math. (-:

George said...

What I usually say is:
I study machine learning, which is about creating programs that improve with experience and extract useful information from data.

You can think of me as a modern-day soothsayer.

Almost always this starts conversations where I mention many applications of machine learning.

David said...

I'm involved in the creation of our eventual overlords.

jed said...

I get computers to learn about things like [everyday example of your specific topic description below].

In my experience, for any statistical modeling worth doing, there's always some example people can understand where it would help (if it worked well enough). Unfortunately I don't know enough about where standard techniques fail in your models to offer an example; if you don't you need to come up with one anyway for other reasons.

So give us a compelling example where we'd like to have really great "MAP inference in graphical models, involving classes of energy functions where standard techniques like quadratic pseudo boolean programming or max-product belief propagation are inefficient or don't work well."

Danny Tarlow said...

Haha, these are great. George, I love it:

"Sooo... what do you do?"

"You can think of me as a modern-day soothsayer."

Captain Wasabi said...

I am building my wife.

Craig said...

I tend to go with:

"I write programs that can learn to make predictions from data"

Clint Laskowski said...

I am an evil genius, that's all you need to know.

tedivm said...

I'm building skynet.

sleepy said...

I write algorithms that teach Netflix to recommended "Weekend at Bernies" to you at some point.

almost_an_understudy said...

@Sleepy is my favorite so far. This whole post is awesome by the way, well done.

stompchicken said...

either...

"I use computers to analyse the past and make predictions about the future"

or...

"Well, I'm not supposed to talk about it but... [lean in and whisper]... I make killer robots"

Дмитрий said...

I am building robots for a new human-machine society.

Дмитрий said...

I am building robots for a new human-machine society.

Benjamin Higgins said...

You do the opposite of encryption.

Justin said...

"I'm building oracles. Yeah, like the drugged girl in 300."

Ivansky said...

what google translate does

Jasper said...

I actually ran into this problem today at the US customs line. The customs agent was running through his usual questions:

Customs agent: 'What are you doing in Canada?'

Me: 'I'm doing my phd'

Customs agent: 'What are you doing your phd in?'

Me: 'Artificial Intelligence'

Customs agent: Laughs and says 'what can you do with that???' then tells me to look in the camera as he takes a photo of my face and scans my fingerprints...'

Me: 'A lot of things...' and restrained myself just enough to get into how ironic it was that he was asking me that question just as he was using two great examples of the utility of AI.

Onur said...

I make computers understand the meaning of observations (or data).