Thursday, January 28, 2016

Mark Zuckerberg Understands the Problem with DeepMind's Brand of AI

DeepMind's GO Playing Program

In the wake of DeepMind's announcement that their GO playing program has defeated the European GO champion, I thought I would write a short post to express my usual rebellious point of view. DeepMind is claiming that this is another giant step toward true AI but this could not be further from the truth. In spite of all the hoopla and the back slapping, I can assure everyone that there has been no breakthrough and no giant step toward true general machine intelligence. This time though, I will let FaceBook's Mark Zuckerberg explain why DeepMind's GO program is not even close to true AI.

Zuckerberg Nails It

In a recent blog post, he had this to say about supervised learning, the kind of machine learning used by DeepMind (emphasis mine):
Our best guess at how to teach an AI common sense is through a method called unsupervised learning. My example of supervised learning above was showing a picture book to a child and telling them the names of everything they see. Unsupervised learning would be giving them a book and letting them figure out what to do with it. They could pick it up and by touching it learn to turn the pages. Or they could let go of it and realize it falls to the ground.

Unsupervised learning is learning how the world works by observing and trying things out rather than being told what to do. This is how most animals learn. It's key to building systems with human-like common sense because it doesn't require a person to teach it everything they know. It gives the machine the ability to anticipate what may happen in the future and predict the effect of an action. It could help us build machines that can hold conversations or plan complex sequences of actions -- necessary components for any authentic Jarvis.

Unsupervised learning is a long term focus of our AI research team at Facebook, and it remains an important challenge for the whole AI research community.

Since no one understands how general unsupervised learning actually works, we're quite a ways off from building the general AIs you see in movies. Some people claim this is just a matter of getting more computing power -- and that as Moore's law continues and computing becomes cheaper we'll naturally have AIs that surpass human intelligence. This is incorrect. We fundamentally do not understand how general learning works. This is an unsolved problem -- maybe the most important problem of this century or even millennium. Until we solve this problem, throwing all the machine power in the world at it cannot create an AI that can do everything a person can.
Zuckerberg is right.

See Also:

Why Deep Learning Is a Hindrance to Progress Toward True AI


Guichong Li said...

It is really interesting if Zuckerberg finds a solution for ML...

Louis Savain said...

Hi Guichong,

Zuckerberg is a very smart man but I'm afraid he's not that smart.

Fergal Byrne said...

I agree 100% with both your premise and Zuckerberg's criticism. Even better, David Silver himself explains in the 7-minute Nature ad (at 5:45) how this really works: it plays millions and millions of games against itself until it learns how to guess (estimate the Q function for) the values of board positions better than P2-level Go professionals. How is this a "step towards AGI" as Demis suggests 45 seconds later?

Louis Savain said...


Thanks for commenting. What I don't understand about Hassabis is that he has supposedly obtained a degree in neuroscience and yet, his approach to AI has very little to do with the brain. Anybody who studies biological sensors, such as the retina, can infer that the brain is inherently a massive timing mechanism. I also fail to understand why Google was willing to pay half a billion dollars for a deep learning outfit. I predict it will turn out to be one of the worst AI investments in history.

keghn feem said...

Very interesting.

John Common said...

read Aristotle. machines will never be intelligent because they do not have what Aristotle calls nous or active intelligence, I.e. that immaterial faculty which abstracts universal concepts from sense data. machines will never have concepts.
the fundamental act of intelligence is the intuitive grasp of being or existence, machines will never perceive being

Eran EranG said...

So, if I can invest a computer program that can play no just GO, but ANY board game, then I should hope to sell if for $5B? Maybe add Monopoly and Chutes and Ladders.

Anyone want to help me get this going?

Anonymous said...

Unfortunately, the arguments mentioned by Louis are one sided.

Some facts ....

Hassabis uses both supervised & unsupervised learning algorithms.
The brain uses both supervised & unsupervised algorithms.
Supervised algorithms produce faster results in nature & in computer systems.

I don't mind debates about what the optimal leaning algorithms are, in fact these will change as computer languages evolve and are incorporated into the evolving hardware.

As we approach the limits of Moore's law eventually a new paradigm in computer hardware will be developed requiring new software to be written which will hopefully be more efficient, approaching that of the brain.

This article & language used reminds me of Savain's previous nonsense about the illusion of space. I appreciate the need to discuss the difficulties that modern science (esp quantum mechanics) has in explaining rationally various observations but when you add ridiculous claims - like there is no space or Hassabis only believes in supervised learning you only belittle your good arguments & people lose interest in the remaining valid arguments you do have.