Machines That Teach Each Other Could Be Decisive For Artificial Intelligence - Alternative View

Machines That Teach Each Other Could Be Decisive For Artificial Intelligence - Alternative View
Machines That Teach Each Other Could Be Decisive For Artificial Intelligence - Alternative View

Video: Machines That Teach Each Other Could Be Decisive For Artificial Intelligence - Alternative View

Video: Machines That Teach Each Other Could Be Decisive For Artificial Intelligence - Alternative View
Video: The Rise of the Machines – Why Automation is Different this Time 2024, November
Anonim

During a press conference to announce the autopilot function in the Tesla Model S in October 2015, Tesla CEO Elon Musk said that each driver will become an “expert coach” for each Model S. Each vehicle will be able to improve its autonomy functions. learning from its driver, but more importantly, when one Tesla learns from its driver, that knowledge will be shared among the rest of Tesla vehicles.

Very soon, Model S owners noticed that the vehicle's self-driving functions were gradually improving. In one example, Tesla's were making the wrong early exits on motorways, forcing their owners to manually navigate the vehicle along the correct route. After just a few weeks, owners noted that cars were no longer making premature exits.

“It's amazing that the improvement happened so quickly,” said one Tesla owner.

Intelligent systems like the ones powered by the latest machine learning software aren't just getting smarter: they're getting smarter faster and faster. Understanding the speed at which these systems are evolving can be a particularly difficult part of managing technological progress.

Ray Kurzweil has written extensively about gaps in human understanding, describing the so-called "intuitive linear" view of technological change and the "exponential" rate of change that is happening now. Nearly two decades after writing an important essay he called The Law of Accelerating Return - a theory of evolutionary change describing how the rate of improvement of systems changes over time - related devices began to share knowledge among themselves, accelerating their own improvement.

“I think this is probably the biggest exponential trend in AI,” says Hod Lipson, professor of mechanical engineering and computer science at Columbia University.

“All exponential technologies have different 'exponents' for trends, he adds. "But this one is probably the biggest." In his opinion, this "machine learning" - when devices transfer knowledge to each other (not to be confused with machine learning) - an important step towards accelerating the improvement of such systems.

“Sometimes this is cooperation, for example, when one machine learns from another, as if they have a swarm consciousness. Sometimes it's a leapfrog, like an arms race between two systems playing chess with each other."

Promotional video:

Lipson believes this path of AI development is powerful, in part because it removes the need for training data.

“Data is the fuel of machine learning, but even for machines it is difficult to get some data - it can be risky, slow, expensive or unattainable. In such cases, machines can share their experiences or create synthetic experiences for each other to supplement or replace data. It turns out that this is not such a weak effect - it is essentially self-reinforcement, and exponential at that."

Lipson cites DeepMind's recent breakthrough, AlphaGo Zero, as exemplary training AI without training data. Many are familiar with AlphaGo, a machine learning AI that became the world's best Go player by examining a massive amount of data of millions of games played in Go. AlphaGo Zero was able to beat even him without looking at the training data, just learning the rules of the game and playing with himself. Then he beat the world's best chess software after just eight hours of practice.

Imagine thousands of these AlphaGo Zeroes instantly sharing their acquired knowledge.

And these are not only toys. We are already seeing the powerful impact of the speed at which businesses can improve the performance of their devices. One example is the industrial digital twin technology - a software model of a machine that simulates what happens to equipment. Imagine a machine looking inside itself and showing its image to technicians.

For example, a digital twin steam turbine can measure steam temperature, rotor speed, cold starts and other data to predict failures and alert technicians to avoid costly repairs. Digital twins make these predictions by examining their own performance and also rely on models developed by other steam turbines.

As machines begin to learn in their environment in powerful new ways, their development is accelerated through the exchange of data. The collective intelligence of each steam turbine, scattered across the planet, can accelerate the predictive power of each individual machine. Where there is one car without a driver, there will also be hundreds of other drivers who will teach their cars, imparting knowledge to everyone.

Don't forget that this is just the beginning.

Ilya Khel