Alpha Go Zero is Making AI Sexy Again with Self Learning and It’s Getting Scary

Google is trying to leverage it’s AI-advantage over rivals with a new range of devices, but the new hardware isn’t that sexy, yet. Far more impressive is the new ground in AI it seems Alpha Go Zero has made.

  • In 2016, AlphaGo, a program created by Google’s DeepMind unit, already defeated the best human Go players, by pretty wide margins.
  • Now its successor dubbed AlphaGo Zero, has “taught itself” without any game history data, and beat the original AlphaGo in a mater series by a margin of 100–0.

Not quite technological Singularity stuff yet, but there’s something more along the lines of “self-learning” going on here that’s surprised a lot of people.

Who knew one of the most ancient games on Earth, would be one of the simulations that best depict the progress that is being made by Artificial intelligence.

Starting from Scratch

Here is an AI that didn’t start from human data, this is an important landmark in AI research.

In their paper published in Nature, deep learning has come to a point where a single neural network that has internalized both policy and value has improved efficiency, pretty much showing that it needs 40x less energy than the earlier version as well.

Alpha Go Zero as DeepMind explains on their blog post, is an AI equipped with reinforced learning, where the AI becomes its own teacher.

Let me let that sink in for you just for one minute. The system starts off with a neural network that knows nothing about the rules, or the game of Go. By playing itself letting the neural network with a powerful search algorithm it tunes and updates its prediction algorithm until it’s mastered the subject matter.

Let me remind you, this is a step beyond the AI that was a step beyond the smartest human on the planet at doing this, a grand-master subject expert.

What this Means for the Future AI

  • For tasks where human data is unreliable or simply unavailable, AI will be able to by-pass these limitations and create a self-learning system.
  • Algorithms that teach itself mastery in a given task or towards a given goal could can achieve “superhuman performance” levels relatively quickly.

That AI can start have a tabula rasa start and finally perform displays an advancement in deep learning that puts the world at notice. If it takes one year to make this kind of progress, what’s next?

Superhuman AI is Kind of Already Here

More powerful predictive algorithms based solely on reinforcement learning, without human data, guidance or domain knowledge beyond game rule is now a dream that has some proof of concept. What problems and tasks in the world can this kind of AI make a difference in?

Modeling AI after humans is probably stupid, if you can create a better algorithm, why not do it. As the original AlphaGo program learned by assessing data from thousands of real-life games played by humans and was inferior to the new one that began as a blank slate, it’s basically showing how AI can easily attain a kind of “intelligence” that’s superior to the sum total of all human experience in a given discipline.

Maybe we shouldn’t try to make AI more human. By skipping over human intuition here, AlphaGo Zero was able to develop strategies more suited to its capabilities. That’s likely the secret sauce here, maybe super-intelligence isn’t human at all, maybe it really is the domain of AI.

F*ck Moore’s Law

There’s a point where AI will mean AI facilitates answers to human problems we’ve never thought of and could never arrive at.

You don’t need to wait for an Einstein if you can frame the right problem blank slate for super-human intelligence to find the right answers to.

Think about it, this isn’t just processing speeds or computational power advances, this is the inkling of self-learning AI.

For those of us like myself who fret about the advent of general artificial intelligence, DeepMind is going to unleash this kind of thing on protein folding and biotech, and the world could change in a flash. If AlphaGo Zero is any case study, it already has.

While you are here, why not check out other tags on Medium:

Top | Today | All-time | Tech | Future | Entre | Crypto | Block | Ed-Picks | Biz | VC | Prod | Paywall | Free | Life | Pod | News | Lessons | Journ | Mark | Self | Trav | Write | Inspi | UBI | AI | Robots | FoW | RW | WLB | Econ | DL | AVs | Des | Crea | Startup | Innov |