New research suggests (AI) Spookily strong artificial intelligence systems may run so well because their structure exploits the fundamental laws of the cosmos,
The new findings may ease answer a longstanding puzzle about a class of artificial intelligence that use a strategy called deep learning. This deep knowledge or deep neural network programs, as they’re called, are algorithms that have multiple layers in which lower-level calculations feed into higher ones. Deep neural networks usually perform astonishingly well at solving problems as complex as knocking the world’s best player of the strategy board game Go or matching cat photos, yet know one fully understood why.
Picture Courtesy: www.gevme.com
A physicist at the Massachusetts Institute of Technology and a co-author of the new research, Max Tegmar said ‘It turns out; one cause may be that they are tapping into the unique properties of the physical world,
Tegmark told Live Science, “The laws of physics only present this ‘very special class of problems.’ This tiny fraction of the problems that physics makes us care about and the tiny fraction of problems that neural networks can work are more or less the same,”
Picture Courtesy: www.insuranceandrisk.com.au
AI accomplished a task many people considered impossible: DeepMind, Google’s deep learning AI system, beat the world’s best Go player after winning the European Go champion. The feat amazed the world because the number of potential Go moves passes the number of atoms in the cosmos, and past Go-playing robots played only as well as an average human player.
But even more surprising than DeepMind’s utter rout of its players was how it finished the task.
For instance, DeepMind was not explicitly taught Go plan and was not trained to understand classic sequences of moves. Instead, it simply ‘watched’ millions of games, and then played many more against itself and other players.
Picture Courtesy: top-stocks-to-buy.com
Another long-held puzzle is why these deep networks are so much better than so-called simple ones, which contain as little as one layer,
To understand why this method works, Tegmark and Lin chose to flip the question on its head.
Lin said. ‘Suppose somebody gave you a key. Every lock you try, it seems to open. One might think that the key has some magic properties. But another chance is that all the locks are magical. In the case of neural nets, I assume it’s a bit of both,’
Tegmark said, ‘One possibility could be that the ‘real world’ problems have unique properties because the real world is very special,’
Picture Courtesy: img.wennermedia.com
Take one of the biggest neural-network mysteries: These networks usually take what appear to be computationally hairy problems, like the Go game, and somehow find answers using far fewer calculations than anticipated.
Tegmark said, ‘It turns out that the math applied by neural networks is simplified thanks to some special properties of the universe. The first is that the equations that govern several laws of physics, from quantum mechanics to gravity to special relativity, are basically simple math problems, the equations include variables raised to a low power.
‘What’s more, objects in the cosmos are governed by locality, meaning they are restricted by the speed of light. ,Practically speaking, that means near objects in the universe are more likely to affect each other than things that are far from each other.’
Picture Courtesy: www.exopoliticsindia.in
Many things in the universe also follow what’s called a normal or Gaussian distribution. This is the classic ‘bell curve’ that governs everything from features such as human height to the speed of gas molecules zooming around in the environment.
All of these special features of the cosmos mean that the problems facing neural networks are actually special math puzzles that can be radically simplified.
There are also problems that would be much difficult for neural networks to solve, covering encryption schemes that secure information on the web; such systems just look like random noise.
While the subatomic laws of nature are simplistic, the equations describing a bumblebee flight are especially complicated, while those governing gas molecules persist simply.
Lin said, ‘the point is that some ’emergent’ laws of physics, like those governing an ideal gas, remain pretty simple, whereas some become quite complex. So there is a lot of extra work that needs to be done if one is going to solve in detail why deep learning works so well. I think the paper puts a lot more questions than it answers.’