AI Software Learns To Make AI Software

Megalith

24-bit/48kHz
Joined
Aug 20, 2006
Messages
13,000
Yo dawg—get ready for the end of the world. Researchers have proudly announced that they are taking strides toward software that is advanced enough to basically reproduce and design even more intelligent versions of itself. As MIT, the University of California, the Google Brain research group, and other big-wigs throw out concepts such as “automated machine learning” and “learning to learn,” the assembly line from hell is becoming more and more plausible. Thanks to Crixus for this story.

…leading researchers are finding that they can make software that can learn to do one of the trickiest parts of their own jobs—the task of designing machine-learning software. In one experiment, researchers at the Google Brain artificial intelligence research group had software design a machine-learning system to take a test used to benchmark software that processes language. What it came up with surpassed previously published results from software designed by humans.
 
The trick is to do all this AI research in airgapped black box rooms with manual destruct capability.
 
There is a line in the movie, "The Hunt for Red October", that comes to mind:


"You arrogant ass! You killed US."
 
As long as it's not able to prevent us from pulling the plug. I immediately thought of that "Person of Interest" episode where "The Machine" moved itself.
 
Withhow good they treat the test robots they make i see this working out well for all of us.
 
Oh google....

If you have to tell yourself "Don't be evil.".... chances are good that ou and your software will probably be evil.

Skynet goes live soon.
 
Yo dawg—get ready for the end of the world. Researchers have proudly announced that they are taking strides toward software that is advanced enough to basically reproduce and design even more intelligent versions of itself. As MIT, the University of California, the Google Brain research group, and other big-wigs throw out concepts such as “automated machine learning” and “learning to learn,” the assembly line from hell is becoming more and more plausible. Thanks to Crixus for this story.

…leading researchers are finding that they can make software that can learn to do one of the trickiest parts of their own jobs—the task of designing machine-learning software. In one experiment, researchers at the Google Brain artificial intelligence research group had software design a machine-learning system to take a test used to benchmark software that processes language. What it came up with surpassed previously published results from software designed by humans.

This isn't quite as intelligent as you think. Let's make weather prediction an example:

Let's say I use 5 variables now to train myself on what is and is not a true condition. I might be able to weight the coefficients on each variable to affect the final outcome.

Chance it will rain:
Weight / Variable
20% There's a pressure gradient of 10mBar higher
30% Front coming in from West
15% It is currently raining to the West
15% Relative humidity is 80%
10% Geographic location, and time of year versus average rainfall versus predicted rainfall
10% Historical match based on roughly similar conditions (Feedback loop variable)

This isn't how real weather models work, (They use recursive cell analysis) but it is a "for example"

So we get A1X1 + A2X2 + A3X3 + A4X4 + A5X5 = value. Value > .5 = true or success. Greater values represent greater success rate.

When you make mutiple rows like this with multiple success rates you set up an Ax = b (where A is your rates and b is your success vector) This is where your "training" comes in. You optimize based on the least square error of the matrix.

Lets call this the "Himalayan model"

Now an AI program can adjust the weight parameters by reducing the matrix (Why GPU's are great at tackling these problems with their vector math and parallel processing for doing things like 2x2 determinants.

But lets say there are a whole slew of other variables, like 30 or 40 of them, that are too complex to add to the equation because it makes compute time too long to figure out all the permutations. Reducing matrices exponentially grows in time as the matrix increases linearly in number of weights. So the AI just adds a variable itself, once at a time, does a preliminary analysis and see if it's current prediction model is any more accurate. Thus AI creating a new formula to make itself smarter...but not really...just able to guess better based on more variables. This becomes "Himalayan model version 2"

It's pretty brain dead simple. You should see some of the fun AI stuff I'm playing with. I have machines that can build themselves much faster than their human counterparts can and for less money.
 
Last edited by a moderator:
Indeed digitalGriffin, these algorithms look for local minima. Reminds of solving differential equations where the 1st derivative is 0. What I find interesting is that models to entirely different problems may end up looking similar. These taxonomies might inform us to underlying mechanisms with predictive value. Do you have a website/blog?
 
I wonder if those MIT researchers realize they killed their own jobs. The AI will figure out how to teach AI and then those researchers are gone.
 
I can't wait for the time when machines do it all, and we just spend all that free time trying to overthrow them :D

Vonnegut was so right! Even the engineers were loyal subjects to their AI master.
 
Last edited:
Back
Top