Teaching machines to anticipate human interactions using videos of TV shows | Hacker News Reader

Teaching machines to anticipate human interactions using videos of TV shows

news.mit.edu

34 pointslettergram15 comments

Hmm, not sure this is a good idea. Most TV characters act so little like normal people that any AI trying to learn anything from them would be getting a really distorted picture.

I mean, one of the very examples they give in the article is The Office. Yeah, I see that going quite poorly as a learning tool.

The communication is on a higher level, not in the first order sense of the words. Looks perfect for differential analysis and stochastic gradient descent, from what was said to what was ment. To have taught a bot some humor would be quiet the relieve for all the worried prophets of the AIpocolypse. The show scripts are planned for a high quality of the language, too.

Edit: my first expectation of the results was rather amusingly like yours

This was my first thought. I would hate interacting with an AI that basically acts like a high school student that watches too much TV. Kinda reminds me of when Microsoft's bot started spouting off Nazi-themed comments on Twitter.

I liked Tay's take on reality. I think we could use more of that. /s At the end of the day I believe AI is bound of have mental problems just like humans have mental problems the reason being the world does not really make sense when you look at it. The things which happen to us all don't make sense. How we do many things don't make sense. There is little logical consistency in the world. Learning from the world and the action of other people is a extremely error ridden process and there are millions of possibilities of learning bad ideas or the wrong thing because we are all like that. We make mistakes all the time and they effect other people as well as ourselves.

I don't know if this idea is stupid, but in an area like politics, I wonder if something like this could be useful for detecting subtext. Politics and diplomacy in particular have a sort of universal protocol for interaction that I wonder if it would be easier to read the subtext of human interaction to get some clues about the overall relationship, just through studying how the interaction deviates from the universal protocol. For example I wonder if you could build a map of hot-cold relationships between countries just by watching how their diplomats interact and using methods like this to try and read the subtextual communication happening.

It may be possible, but it would be very hard. Every culture has its own way of interacting and some cultures can have one action that's polite - but in another culture that's completely offensive.

My guess is it wouldn't work for diplomats, but potentially in a large country like the U.S., India, China, it may work for the standard individual or politician.

But it's because of those cultural differences (and the fear of breaking a taboo) that diplomats learn that sort of suit-and-tie-or-thobe, shaking-hands-and-smiling, business-like-but-not-really thing that we are all used to seeing on television. That's what I mean when I say "universal protocol of interaction". Because it is a template with fairly standard rules, I imagine it would be easier to "predict" how two humans will interact like the technique demonstrated in this post. But then beyond that, I figured the interaction of diplomats might be so easy to predict ("smile, shake hands or bow, speak greeting"), that deviation would be more pronounced than for example the scene from The Office. Enough that some useful information might be conveyed in the deviation.

So that's how we are making Skynet hate humanity.

I guess I see a bit more depth in this comment than those who downvoted it. (Yes, it was funny but it made me think.)

Here's a theory: some people think that syrupy positivity is so important in the startup mentality that dissenting voices are tantamount to "griefers" who should be ignored or, if possible, silenced. It's an easy call to make and it only takes one or two votes from those quick on the draw with the downvote button.

Considering what goes for tv shows these days, i can anticipate problems.

However, for a "common sense" grade library that might be adequate with some careful picking.

Looks like a "cliche detector" to me.

No line so unfunny, that future AI generations will not drown it in a thundering applause. <audience dying of laughter>

Neural Network seem to go where the data is. The data (scripts) to TV-Shows are ready available. Real conversations are not (Youtube-Subtitle, anyone?). Shirking hard problems today might help , to train a generation that can understand dialects and muffled voices from context tomorrow.

That is the premise of "Explorers", albeit aliens, not AI.

http://www.imdb.com/title/tt0089114/

Reply on news.ycombinator.com