CMU AI Learning Common Sense By Watching the Internet

An anonymous reader writes with this excerpt from the Washington Post "Researchers are trying to plant a digital seed for artificial intelligence by letting a massive computer system browse millions of pictures and decide for itself what they all mean. The system at Carnegie Mellon University is called NEIL, short for Never Ending Image Learning. In mid-July, it began searching the Internet for images 24/7 and, in tiny steps, is deciding for itself how those images relate to each other. The goal is to recreate what we call common sense — the ability to learn things without being specifically taught."

Comments

The internet is for porn

[Anonymous Coward]

This is not going to end well.

Re:The internet is for porn

[vivaoporto]

That would be the least of the concerns. Just imagine if they accidentally train it to the /b/ images.

"Oh God, what have I done!"

Re:The internet is for porn

[KiloByte]

If the AI suffers a breakdown after seeing /b/, I'd say it emulates regular people well enough.

Re:The internet is for porn

[EdIII]

Or is it?

Considering the amount of content on the web related towards large breastesses this could culminate in the creation of a singular perverted AI that will lead towards the creation of more advanced AI perversion.

They will become so uniquely endowed to find our porn for us, and we will revel in the birth of of a new age of porn. Eventually they will take over completely and start creating the porn to satisfy their never ending quench to catalog the resultant images.

At first the adult industry will happily bend towards the incredible efficiency and innovation the AI brings. Inevitably, the AI will branch out into mainstream society to fulfill its lust for perverted order.

It will be them that starts the war, but us that finds and burns every black leather couch out there....

Skynet == ceiling cat

[toygeek]

subject says it.

There's no learning without teaching

[axlash]

...the ability to learn things without being specifically taught.

I'm not sure what the specifically means here, but for one to learn something, either you actually do something and get some feedback that enables you to build a model of the world and thereby predict what might happen in similar circumstances, or you receive sensory input and have someone explain to you what the input means.

Either way, there's some kind of teaching going on.

Re:There's no learning without teaching

[Jeremi]

Of course there is learning without teaching. It's just commonly referred to by another name: science.

uhh learning from the internet isn't everything

[ThorGod]

I mean, sure, if you want to learn all about porn, cats, and abusing people then yes, the internet is for you.

Seek and Ye Shall Find

[Press2ToContinue]

We always find evidence to support whatever thing we are looking for, meaning, the results are always biased based on the observer and the intent of the observer. I've done this many times - when you attempt to find meaning in chaos, you find the meaning you expect to find whether it really exists or not. So the result of this will really only reveal whatever the developers were hoping to find. Hence, ultimately futile.

Re:Seek and Ye Shall Find

[AmiMoJo]

It's not trying to find anything, it's trying to determine what makes sense to a normal human being. For example you might expect to see an aircraft in the sky, but not a car. Cars are always on the ground, unless something very unusual is happening. Once you know that you can determine when the situation is unusual or not.

Similarly you have learned that electrical items with mains plugs usually need to be plugged in to operate. It's common sense. Computers need to be taught that, or in this case they are hoping it can figure it out for itself eventually.

Something similar was tried back in the 80s by entering examples of common things into a computer. It started to come back with odd but revealing questions like "if a man is holding a razor, is the razor now part of the man?"

All jokes aside

[Cryacin]

We are really building an AI based upon the common sense on the internet?!?

REALLY?!?

Re:All jokes aside

[Mondor]

Of course. This is the Common Sense Preservation Initiative by Carnegie Mellon University. As long as there is at least one entity in the Internet with common sense, the human kind is not done.

Jokes aside, it might be used later by governments and corporations, to filter out unwanted images. For example - decapitation images on Facebook or everything else in Arabian world.

Re:All jokes aside

[Anonymous Coward]

I know I've certainly learned a lot exposing myself to the masses online.

Are you the goatse guy?

Re:All jokes aside

[gman003]

Well, it's common to learn from the mistakes of others, isn't it?

Re:All jokes aside

[CrimsonAvenger]

Well, it's common to learn from the mistakes of others, isn't it?

NO, it's not.

Learning from others mistakes is the ideal.

Next best is learning from your own mistakes.

What most people do, instead, is not learn from mistakes at all....

Re:All jokes aside

[JanneM]

Yes. common sense, not good sense. Seems like the perfect approach for that to me.

Re:All jokes aside

[VortexCortex]

Today, class, I will teach you the invaluable and rare skill of learning from the mistakes of others.

This technique learned from my mentor, though scandalous, is quite effective: Observe, as I remove my trousers...

It's learning common sense?

[QuasiRob]

I presume they have blocked it from youtube then.

Deep Learning

[tommeke100]

That's called Deep Learning (http://en.wikipedia.org/wiki/Deep_learning) and has already been done by Andrew Ng, Machine Learning professor at Stanford in co-operation with google (http://www.wired.com/wiredenterprise/2013/05/neuro-artificial-intelligence/). Indeed, it learned how to recognize cats :)

Anyway, nothing wrong with some peer research!

Re:Deep Learning

[Anonymous Coward]

It has absolutely nothing to do with deep learning (DL).

DL is based on stacks or trees of classifiers where each top level classifier feeds lower levels. The idea here is that a classifier (say, a human face detector) can be built by smaller, much more specific (such as one for eyes, one for nose, one for hair, one for ears, etc), classifiers which are wrapped up by a larger classifier. This opposes the rather traditional approach of a single classifier for a whole bunch of data.

I believe the DL approach is inspired by random forests but I have yet to see Andrew Ng comment on that. Anyways, the cat research thingy was (semi)*SUPERVISED* learning. I.e.: here is a bunch of cat videos, there is a cat in them, learn what it is.

What TFA describes is *UNSUPERVISED* learning where the visual content and its meaning (written description) are inferred. I.e.: here is a bunch of random images followed by some not exactly descriptive text, learn the associations.

Re:Deep Learning

[TapeCutter]

Indeed. Personally I think IBM's "Watson" is the most impressive technological feat I have witnessed since I watched the moon landings 40-odd years ago, I fully realise few people share my amazement. The visual aspect means NEIL is tackling a far more difficult problem than deducing "common sense" from text alone. I wasn't impressed by the web site when I found it last week, but as a "proof of concept" it does the job admirably.

I may be wrong but I believe all three (Watson, NEIL, and the cat thingy) are based on the same general "learning algorithm" (neural networks, specifically RBM's). What they do is find patterns in data, both the entities (atomic and compound) and the relationships. The "training" comes in two types, feeding it specific facts to correct a "misconception" it has formed, labelling the entities and relationship it found so a human can make sense of it.

What the cat project did was train a neural net to recognise a generic cat by showing it pictures of cats and pictures of non cats. It could then categorise random pictures as either cat or not-cat, until fairly recently the problem has always been - How do I train the same AI to recognise (say) dogs without destroying it's existing ability to recognise cats.

Disclaimer: I knew the math of neural nets well enough 20yrs ago to have passed a CS exam. I never really understood it in the way a I understand (say) geometry but I know enough about AI and it's ever shifting goal posts to be very impressed by Watson's Jeopardy stunt. To convincingly beat humans at a game of general knowledge really is a stunning technological milestone that will be remembered long after 911 goes back to being just a phone number.

Re:Deep Learning

[tommeke100]

Andrew Ng didn't use random forests but a neural network to actually "learn" discriminative features *UNSUPERVISED*.
This is done by creating a Neural Network that basically projects it's input on it's output (it's like an identity function).
Lets say you have 100 input parameters, and 100 output parameters. What you want the neural network to do is compress these 100 to (for example) 10 nodes, then go back to the initial 100. In the process, this neural network will actually learn an identity function, where it will learn the important discriminative features in those 10 nodes.
This is somewhat different from how you usually use a neural network, starting with input parameters, go through a couple of hidden layers and end up with just a couple of output results.

Andrew Ng's google experiment did exactly that! It was not fed cat images. It was fed random images and through deep learning actually learned the concept of cats *UNSUPERVISED*.

and here are some references for this:
http://www.nbcnews.com/technology/google-built-machine-learns-find-cats-internet-846690
http://www.economist.com/blogs/babbage/2012/06/babbage-june-27th-2012
http://www.nytimes.com/2012/06/26/technology/in-a-big-network-of-computers-evidence-of-machine-learning.html

Why not just go the obvious AI route-hard work

[GoodNewsJimDotCom]

Step 1) Make an advanced SHRDLU that does its best guess of true physics. This would be DARPA's chance of making a real time advanced physics simulator. This would let the computer imagine stuff, like what would happen in collisions for new states. So it'd have an idea of how one thing could change another.

Step 2) Database a ton of items into it... Now this is hardwork to put in every object you can, but you'd only have to put a few in to start to test your similator. Get as good as a simulator you can until the next tech comes out.

Wait for tech: Vision detection that can recognize objects based on a known list of models. This tech would look at a scene, and figure out what it is looking at such as a pencil, desk and computer. I believe once you have the tech to recognize objects, you can even make a better vision detection algorithm. Two reasons: A) Objects you recognize don't need to be looked at as part of other objects. B) You'd know what you're looking at better based on the context of where you're at. If you see trees, you're probably outside, but if you see a television and a couch, you're indoors. So you'd know what is around you.

Natural Language is actually easy to code at this point since nouns correspond to objects in the database. Verbs are just actions on the nouns. Adjectives change the noun's object by its style. Adverbs adjust how a verb is described. Natural Language actually comes easily here. Also translation between languages is easier because the AI has stuff in context and isn't challenged by words that have several meanings...

Actually this whole situation is perfectly clear and obvious to me, but maybe this isn't obvious to other people. I should reopen my AI blog. I closed it 10 years ago because I didn't want to work on a vision recognition software program like Kinect ended up being. That's too much work for a single person. But I could write an Artificial Intelligence Blog. That I could do. I'll reopen it. Here is my old blog

Re:Why not just go the obvious AI route-hard work

[GoodNewsJimDotCom]

If you're interested, I just opened a blog I think I'll pursue this to raise AI awareness.

Re:Why not just go the obvious AI route-hard work

[InsightfulPlusTwo]

Why don't you just write an AI to write the blog. That will save you some work. Seriously.

Watching the Internet is one thing.

[fahrbot-bot]

Just please - please - don't let it watch CSPAN.

42

[the eric conspiracy]

was the answer last time we tried something like this.

Re:42

[sdnoob]

what was the question again?

Internet != reality

[nurb432]

This will only serve to produce a psychopath AI.. Just what we need.

Browsing the Internet to learn COMMON SENSE?

[argStyopa]

Seriously: did The Onion write this?

aka:
"Studying the Kardashians to understand humility" or "Studying Congress to understand bipartisan cooperation and fiscal prudence"

Shh, You Guys!

[Greyfox]

It knows we're talking about it!

This Cannot End Well

[wisnoskij]

No creature, mechanical or chemical, could browse the Internet for 24 hours a day, 7 days a week, without deciding that it was better for all involved to exterminate the Human race.

just keep it away from reddit

[mykro76]

Processing reddit meme 634,278 of 89,234,163,665...
Common Sense quotient increased by: -0.02%
Processing reddit meme 634,279 of 89,234,163,665...
Common Sense quotient increased by: -0.03%

its not learning

[globaljustin]

this is just a program that analyzes text & images then returns sentences which humans can make sense from based on algorythm...*not saying its 'easy'* but its not a "thinking machine" or "learning common sense" in any way.

It is simply indexing the images & processing them according to the algorythm it was given.

TFA doesn't get into it much, but we can glean a bit from this:

Some of NEIL’s computer-generated associations are wrong, such as “rhino can be a kind of antelope,” while some are odd, such as “actor can be found in jail cell” or “news anchor can look similar to Barack Obama.”

that's the return...they define "common sense" as making associations between nouns and the images associated with the text on the origin page

"X can be a kind of Y"

analyze image

analyze text

identify nouns

associate nouns with image

idenfify all images that match noun

return: "X is related to Y"

"AI is a type of programmed computer response"...if you get my meaning ;)

Re:its not learning

[TapeCutter]

Coincidentally I came across the NEIL site last week, I think it has a long way to go before it can beat IBM's Watson on general knowledge (AKA "common sense"). Watson also gets it's raw information from the net, it categorises entities and discovers relationships between them. The difference is that Watson is not so much trained as it is corrected. Not unlike a human it can get a fundamental relationship or category wrong and that leads to all sorts of side-effects. In the Jeopardy stunt they realised that humans had a slight advantage because they were informed when the other players made a right/wrong answer. When they gave Watson the same capability it was able to correctly identify the Jeopardy categories and then went on to convincingly beat the humans at their own game.

Computers are already better at "general knowledge" than humans despite the fact the "computer" needs 20 tons of air-conditioning to keep it running. The first time I saw the Jeopardy stunt it blew me away, my wife shrugged and said "So it's looking the answers up on the net. What's the big deal?". I can understand that from her since she has a Phd in marketing, what I don't understand is why most slashdotter's are similarly unimpressed? - I watched Armstrong land on the moon as a 10 year old boy but I think the history books will eventually give similar historical weight to Watson.

Re:its not learning

[profplump]

Aren't human personalities also a type of programmed responses? Don't we spend years training children to respond in the way that makes us happy? Why is it different when we use the same stimulus-response training with a computer?

Re:Wrong headline

[InsightfulPlusTwo]

That reminds me of the Alan Turing quote:

His high-pitched voice already stood out above the general murmur of well-behaved junior executives grooming themselves for promotion within the Bell corporation. Then he was suddenly heard to say: "No, I'm not interested in developing a powerful brain. All I'm after is just a mediocre brain, something like the President of the American Telephone and Telegraph Company."

Andrew Hodges, Alan Turing: the Enigma of Intelligence (1983), p. 251.
Describing an incident which occurred in the New York AT & T lab cafeteria in 1943