Kevin Schofield - Tour of Microsoft Research, Part II (machine learning)

TDude wrote:

Can a machine learn? I thought it can only be programmed.

Can the sound barrier be broken?  Can we land on the moon?  Can we colonise Mars?  Can machines learn?

Just one of those things that is impossible until someone does it.

<drools while waiting for video download>

It's just weird that everyone speaks about "learning" a pc something, while it's only doing what it's programmed to do. It's not like a pc can interact with you directly, even when sometimes you think it's doing that.

Shaded wrote:

<drools while waiting for video download>

Smalband eey? Try with WMP10, you can stream it without downloading it first.

Actually, I thought that was me talking about demos. I love a good demonstration of what someone is working on.

As to whether machines can learn? Certainly not the way we learn, but they can be trained to look for patterns and present those to us in new ways to help us out.

For instance, it can watch the patterns of how I read news feeds. Look at my linkblog at http://www.scobleizer.com for instance. I read 3500 items and put only 100 items a day onto that. Could an algorithm be crafted to look for patterns that I use? How accurate could it get?

I'd love to find out.

scobleizer wrote:

Actually, I thought that was me talking about demos.

Sorry, can't hear the difference on smalband

scobleizer wrote:

For instance, it can watch the patterns of how I read news feeds. Look at my linkblog at http://www.scobleizer.com/" target="_blank">http://www.scobleizer.com for instance. [...] Could an algorithm be crafted to look for patterns that I use? How accurate could it get?

I'd love to find out.

The answer is already on your website: PodCasting! (Future of Podcasting | from: Reel Reviews | MWGblog)
In three words: spoken news feeds.

Charles: wait until you see tomorrow's video. We meet a researcher working to cure HIV (and the folks who did a cool traffic app for the SmartPhone and the technology inside MSN's desktop search).

TDude: that might help, but won't solve the problem.

Here's one: how will the Scobleizer live on and keep working even after my death?

Immortality here we come! Have the computer learn my patterns and it can keep doing what I'm doing even if I retire or die.

Machines can learn things just fine.  Yes, the domain of knowledge is restricted, but this is due to the inherent brittleness of current programming and development paradigms, not because machine learning is inherently impossible.

I also think the assertion that "because computers are programmed, anything they can 'learn' isnt learning" is flawed argument.

The same goes for machine intelligence.

[p.s. This was a very motivating video]

Machines can't learn.

They can see patterns in what they do, but you have to tell it what to look for in patterns, then what to do with it.

If I had a microphone input, my form1 wouldn't start doing text to speech unless i told it i was speaking, and these words represent strings or actions, which we pickup as children.

And as children, we can't be told that they are actions, we learn. The day microsoft releases an application that can learn without being told what to learn, is the day computers learn.

Then again, Scoble seems to learn... odd..

Machine Learning != Human Learning

But machine learning agents can be 'taught' to 'perceive' events and attain an 'awareness' of their environments.

There is no magic here. We're not talking about evolving consciousness and an era of spiritual and ethical 'beings'. Rather, its a rigorous application of classical statistics and probability, with some novel math tricks. 

For those proficient in C# who wish to apply these concepts for themeselves there is a .NET library for creating reasoning algorithms based on Support Vector Machines (SVM).

Link:
http://portals.mit.edu/DevShell/DesktopDefault.aspx?tabindex=5&tabid=1125

ok,
aq

"Can machines learn"?  Of course they can.  They can learn better than humans.  Once you teach them something they remember it forever.

Look at Wikipedia - there's a huge amount of knowledge there.

Does the server really know the information?  Who cares?  All that's important is it can parrot back the information if we ask for it correctly.  How else can you judge knowledge?

See John Searle's Chinese Room analogy.

Two examples of machine learning I'm most familiar with are in games.

One is "Black and White" in which you are a god, and you possess a creature which can be rewarded or punished when it performs actions, and thereby learns what kinds of actions it is and isn't allowed to perform.

A better one is the Creatures series of games. There's a free version of this you can download called "Creatures Docking Station". This simulates artificial creatures, possessing both a simulated brain and simulated genetics. The creatures are capable of learning from their experiences, e.g. food gets rid of hunger, sleep cures tiredness, a good way to not feel crowded is the move away from other creatures, etc. A very interesting phenomena is that there's a machine which teaches them vocabulary, but has the side effect of making them very happy. If left unchecked, an entire population can end up addicted to the machine, using it over and over again. And the problem is they tell each other that the machine cures sadness.

So yes, machines can definatly learn. The difficultly is in making a system that learns successfully. It's very possible that a machine will learn stupid things from the user if the user does stupid things.

Epsi: wait until you see the traffic app on tomorrow's video. There the machine learns from stupid traffic behaviors. Like how we all go home at 5:30 p.m. and clog the freeway. It is quite good at predicting when the clog will clear, as you'll see tomorrow.

The Channel 9 Team wrote:

Getting a glimpse into Eric's office is like getting a glimpse into the future.

Interesting glimpse.  Still appears to be an expert system if it can't take the theories of one area and apply it to another.

But what do I know.  I don't have a college degree.

...neither did Bill Gates.

... or the Wright brothers.

It sure will be a fun race to win.  Stay tuned.

If machine could learn & be independent, you MS-research guys surely have thought about if a machine can become sentient. Is there a sentient test? I don't think the Turing test is enough.

Should we head off the eventual subjugation of the human race and abandon this line of research? Maybe Hollywood is wiser than we think.

Minh wrote:

If machine could learn & be independent, you MS-research guys surely have thought about if a machine can become sentient. Is there a sentient test? I don't think the Turing test is enough.

Should we head off the eventual subjugation of the human race and abandon this line of research? Maybe Hollywood is wiser than we think.

Or they are just not very creative.

The part that gets me about all the Hollywood movies about the machines taking over is....


... if you were a machine....

and you could issue a command say.... sleep(years=70,000)

Why would you be in a hurry to wipe out humans?

I think self repairing machines would be the ultimate procrastinators beyond any homer simpson we've ever seen simply because there would always be a tomorrow.

If anything a machine that didn't like humans would say "see ya later" and go colonize the asteroid belt and replicate to a heathy sustainable population and stay away from the uncertenty of human interaction.

... wouldnt THAT be the ultimate insult?  <Grin>

Shaded wrote:

I think self repairing machines would be the ultimate procrastinators beyond any homer simpson we've ever seen simply because there would always be a tomorrow.

Unless you allow the machine a sense of self-preservation & the ability to calculate & identify threats. Now, you can say let's program the cyborg w/ the 3 laws of robotic, but you know the military would be the first user of these cyborg, so they'd have to tweak the "I shall not harm a human" rule. Other cyborgs could learn to ignore that rule, too. Then get ready to black out the sky, 'cuz here come the cyborgs.

scobleizer wrote:

Epsi: wait until you see the traffic app on tomorrow's video. There the machine learns from stupid traffic behaviors. Like how we all go home at 5:30 p.m. and clog the freeway. It is quite good at predicting when the clog will clear, as you'll see tomorrow.

Hey, didn't SimCity do this like a decade ago?

The guy is both freakish (his stare especially is unnerving) and also brilliant.  Just hearing him talk gave me lots of ideas to simplify a user's experience.  A desktop search is only the first step of organizing simply the information we have...

That priority email thing is cool - is it commercially available in some form? 

Dan

Nope, not yet. I don't even have it yet. I'll ask Kevin if it can be released on the Research site, though.

I don't care if the software can pass a Turing test or not. It's tough not to anthropomorphize here, but I would be happy if it could understand and be 'aware' of when I am busy, and divert calls or messages 'intuitively.'  It would agregate behaviour as Eric Horvitz said. I think the draw back in the selling of this is that your boss could hack your assistant - so to speak, and find out how productive you are. Or are not. And performance reviews could be contingent on your assistant's word you were working or not. I would still like to see one in a couple of years when it is more widely available. Microsoft SideKick or Microsoft Gatekeeper have a nice ring to it. Perhaps Gateskeeper?

How many of the black cubes does Eric Horvitz have? It looked like a obsidian Rubic's cube farm on those shelves.

I was really impressed by the priority program he was using to sort his emails? Anyone know if there's an existing app like that today for outlook? I could really use this kind of app to get through the heeps of mails I get every day

Epsi wrote:

It's very possible that a machine will learn stupid things from the user if the user does stupid things.

And the virus problem will continue. Can you imagine that? Computers executing virusses themself. Virtual suicide.

This is a great video. I love to see what's possible today using advanced computer systems. This reminds me of another research project at MIT I have read about some time ago.

http://web.media.mit.edu/~nathan/research/overhear/

To me, it seem like those two things combined could really have an impact on how we can organize our information (and essentially our life) by the help of "software agents".

Can a machine learn?

I cannot think of a way in which a human being is not a very complicated machine.

People are deluding themselves if they think their mental states are not driven by physical laws, and that the their ability to learn is a result of the design or structure of the atoms that constitute them.

We obviously have several orders of magnitude to go before we get to human level learning, but it is obviously possible -- we have a prototype. 

The Channel 9 Team wrote:

(he is responsible for more patents at Microsoft than any other person)

Has he patented ones and zeroes?
/obvious

Where can I get that email prioritizing program?

(Edited.)
Strange. I was having problems downloading the video, (it was going very slow and I couldn't get the whole thing). Now it's working fine, and humming along at 230 kB/s.

Very cool video!

Hey Robert,

It would be cool if Channel 9 could use Eric's "capture the most memorable moment" or "select the best picture" technology to select the image used to highlight the video. It would probably save you the headache of looking thru 20+ minutes of video for the "perfect shot"

Machines can't learn.

They can see patterns in what they do, but you have to tell it what to look for in patterns, then what to do with it.

If I had a microphone input, my form1 wouldn't start doing text to speech unless i told it i was speaking, and these words represent strings or actions, which we pickup as children.

And as children, we can't be told that they are actions, we learn. The day microsoft releases an application that can learn without being told what to learn, is the day computers learn.

-----

The issue of what action to take after something is learned is not, strictly speaking, part of machine learning. As for whether computers can learn without being told what to learn, I think the answer is both yes and no. Yes in the sense of unsupervised learning techniques such as clustering (K-means) that discover clusters within data via squared-distance minimization. No in the sense that even unsupervised learning techniques have various parameters that need to be adjusted to give reasonable solutions e.g. in K-means one needs to specify the number of clusters. Incidentally, brain research has shown that a form of unsupervised learning -- competitive learning -- occurs when we learn about smells (olfaction). One application of this form of learning has been for object recognition technology.