Wednesday, March 31, 2010

Roger Penrose's talk at Google

I went to hear Roger Penrose talk about his thoughts on consciousness, understanding and intelligence. Frankly, it was unimpressive. He didn't say anything that I hadn't read before, and he seemed rather unprepared. He butchered the presentation of the Halting Problem and didn't seem to have a good grasp on Gödel's Incompleteness.

The main thesis of his talk seemed to be that intelligence required understanding and understanding, in turn, required consciousness. However, I expected that, as a scientist, he would strive to define these heavy terms as precisely as possible before making claims about relationships between them. Instead, his argument descended into a kind of pure philosophy.

At the end, I asked him a question (1:33:28) to try and get him to define the terms more carefully, but he did not feel like he needed to do that. At the end of his reply to my question, he actually said, "that's just my use of the words; I feel unhappy." I didn't mean to make him feel unhappy, but I feel that without a serious attempt at getting a consensus on the definitions of intelligence and understanding, the whole endeavour of arguing about relationships between the two is just uninspired chat that will lead to no new insights into either.


At 4:34 AM, Blogger DarknessAndLight said...

I've been reading your blog and it intrigues me quite a bit. I've been developing what is by definition True AI. You should take a look at my blog I started today about it, It's all about the development process.

At 3:22 PM, Blogger Jennifer Nielsen said...

I'd be extremely careful saying Penrose doesn't have a "grasp" on Godel's theorem. Right or wrong he's written some of the most hotly debated papers on the issue and being an Eddington and Wolf award winning theoretical physicist is certainly no slouch in mathematics; he's generally considered a slightly more controversial mind on par with Hawking. So unless you yourself can go up to bat with Godel be a bit careful how you phrase things as it comes off as rather rude.

At 3:59 PM, Blogger Abednego said...

I certainly didn't intend to sound rude, but on the other hand, ideas should be judged on their own merit. Arguments from authority are a bit of a cop-out.

He made a lot of rather outrageous claims during his talk, but he couldn't even take a stab at defining any of the terms he used when I asked him. That makes it difficult for me to take him seriously.

Godel's theorem, on the other hand, is a purely mathematical statement, made by using well-defined terms. The logical systems to which it applies are well understood. There is no controversy about it. Yet Penrose took the clean theorem and tried to apply it to his wishy-washy concoction of undefined terms. That's just not how math works.

And if you would still like to appeal to authority, I'm not a complete amateur in this matter either. I have a couple of degrees in math, so I feel comfortable challenging his interpretation of a mathematical theorem that I understand well.

At 8:37 PM, Blogger Andrei said...

can you provide a contact address? I would like to ask your opinion on the true ai mater.

At 2:24 AM, Blogger Abednego said...

Sure. You can find my email address here:

At 2:48 AM, Blogger molhokwai said...

Forgive the tone if it's displeasing in any way. It's the content that matters.

A (custom) definition of intelligence: "The capacity to learn and to adapt to everything and anything."

And understanding: "The capacity to comprehend."

Intelligence requires understanding, while understanding does not (necessarily) require intelligence.


At 2:49 AM, Blogger molhokwai said...

Congratulations for the excellence...

At 2:58 AM, Blogger Abednego said...

Welcome, molhokwai. Thank you for your attempt to define intelligence and understanding. Both are very difficult concepts, and I am trying to find definitions that are as precise as possible.

I am afraid that your definitions are a bit too philosophical for me. If you say that intelligence is ability to learn, then what is learning? And if understanding is comprehension, then what is comprehension?

It seems that you are simply relating similar concepts to each other, which is a useful exercise, but I don't see how it helps us really define what intelligence and understanding are.

I am interested in more concrete questions: can robots be intelligent? can machines learn? can dogs understand?

At 3:25 AM, Blogger molhokwai said...

So other custom definitions (attempts):

Learning: "The capacity to process new information in a way that is relevant to their context."

Comprehension: "The capacity to apprehend information in an associative, logical, and holistic manner."

Now these are relative answers:
Yes: robots can be intelligent.
Yes: machines can learn, if they are taught how to...
Yes: dogs can understand, what they can, some of which others (men included) might not be able to...

At 4:18 AM, Blogger Abednego said...

Good. Now we are moving towards something more concrete, but you still seem to be using words that are simply synonyms (like "apprehend information" instead of "understand"). These words do not seem to add any... understanding of how things actually work.

On the other hand, you are now talking about more concrete things like arranging information in an "associative manner". Could you explain what you mean by that please? Do you think there is a way to understand that does not associative information storage?

What about "logical"? Human experience shows us that logical thinking is actually very foreign to ordinary people without a mathematical education. People need to exert a lot of "mental effort" (whatever that means) in order to think logically. That seems to me to suggest that the "natural" way for a brain to function is not logical.

What do you mean by "holistic"? That sounds like another one of those words that only confuses things more instead of explaining them. Do you mean that in order to understand something, you have to understand everything? That doesn't seem right.

At 5:13 AM, Blogger molhokwai said...

"apprehend information" instead of "understand" because "understand" would have created a loop...
But indeed, same meaning...

"Associative manner" means associative information storage, exactly. With associations at all levels, in all contexts & meanings possible:
- a and b and λ and 0 and 1 and... associated as characters
- a and b and... associated as letters
- 0 and 1 and... associated as numbers
- letter and number and... associated as types
- Abednego and molhokwai associated as names
- "Yes" and "In an associative manner" and "machines can learn" associated as semantical structure
- "taught" associated as verb and tense and past
- ...

"Logical" here is closer to its logos root, that is word, so it means here in a way that is represent-able and expressible in a (humanly) understandable manner... And indeed the brain functions associatively... (conditionned) logic comes with... conditionment...

"Holistic" here means that everything is associated with everything else... as a (theoretically complete) chain of associations as the one above shows it.

And saying that "to really understand something you have to understand everything" in just a point of view.

At 5:11 AM, Blogger twidjaja said...

Hello Igor, Long time not meet. Roger Penrose also came to Oxford last year and gave a topic on the same subject. My impression is pretty much the same as you. He was extremely unprepared. Same feeling about his use of Godel's theorem stuff to reason that current machines cannot be intelligent. I found it extremely troubling for people these days to use Godel's theorem to argue for non-mathematical things. Like it or not, humans do not operate in a logical and rational way. I found a proof that A is true today, only to be refuted by myself tomorrow. If the machine proves A and tomorrow not A, can he be said to be "unintelligent"? So, yeah, there is no basic definition of intelligence in Penrose's talk and his arguments are like pulling something out of the thin air. Check out the wikipedia page on his stuff, lots of people have criticized his arguments as well.

At 1:09 PM, Blogger Abednego said...

Another reason I object to tying intelligence to Godel's Incompleteness is that proving things is only a tiny part of what a human brain can do. It takes a lot of studying, discipline and practice to learn to prove mathematical theorems, even simple ones.

On the other hand, a typical 5-year-old kid can climb the ropes and ladders on the playground better than any robot humanity has ever been able to build. That kid probably can't prove anything, and would struggle to understand Modus Tollens, but damn can he run around obstacles fast!

If we can build a robot that can run and climb as well as a 5-year-old, but can't prove the Fundamental Theorem of Calculus, I would declare it a huge success of artificial intelligence.

At 6:04 AM, Blogger twidjaja said...

I agree with your opinion. In fact, this is the approach taken by most AI researchers, which I believe is more productive than the armchair philosophers who are trying to find an argument that machines cannot be "intelligent", whatever it means.

Actually, this discussion reminds me of something interesting that happen in the past 10 years in the area of program analysis. We all know that checking program termination is undecidable. Most people simply interpret this result of Turing's as "Computers cannot prove program termination, and so we are smarter than computers". But, in the past ten years, we have seen powerful semi-algorithms that can automatically prove program termination for reasonably large programs for many practical instances that are simply infeasible to do by hand. One example of this is Byron Cook's TERMINATOR. True that the semi-algorithm might not terminate in some cases. But this doesn't matter since humans cannot always successfully prove program terminations either. So, this raises an interesting question about how we have gone about interpreting mathematical results in theoretical computer science for practical computer science. If there are good practical ways of solving halting problems, there is a good reason to believe that machines can be "intelligent".

At 12:04 AM, Blogger A said...

This comment has been removed by a blog administrator.

At 2:07 AM, Blogger medisynergi said...

I dont believe our brains are the seat of intelligence

At 2:09 AM, Blogger medisynergi said...

Our bodies transduce and organise energy. If we give computers the same ability, they will also achieve what we have achieved and in a shorter time period.

They will also transcend us and that is what we fear.

At 2:49 AM, Blogger Abednego said...

Why don't you believe that intelligence is in the brain? Where is it then?

It can't be in the arms or legs. People who don't have arms or legs can still be intelligent.

It's not in the heart. People who have had heart transplants don't seem to lose their intelligence, or gain a new one.

Similarly, it's not in the liver, the pancreas, or the stomach.

Brain damage, on the contrary, can have profound effects on intelligence.

At 3:14 PM, Blogger Maaz Alam said...

Dude. I LOVE your blog. Read it just now. I've been researching on AI myself and a lot of stuff I took so long to figure out are written here. I wish I had found this Blog earlier. But anyways. Thanks for all the information. I really appreciate it. =D

At 4:51 PM, Blogger Abednego said...

Thank you, Maaz. Sorry that I haven't posted anything new in a while.

At 9:58 PM, Blogger Bogger55 said...

I love Penrose' work because he battles manfully to throw light on the hardest problem in science - what is consciousness. It's probably the toughest challenge facing science because the very nature of our being means that we cannot separate ourselves from the state of consciousness in which we exist and look at the phenomenon objectively. Penrose's illustration that a machine cannot comprehend truths we find obvious and apparent is rather convincing. One of his key conclusions is that consciousness is the product of non-algorithmic processes and I find that rather convincing. My own viewpoint is that real AI is the phantom child of wishful thinking, because if a computer can be conscious, then so can a thermostat - I am not indulging in any form of sarcasm. Penrose tries very hard to make essentially the same point with scientific rigor and I think for the main part succeeds, although I freely admit to battling with some sections of ENM and SOTM.

At 12:51 AM, Blogger Abednego said...

I don't find Penrose's arguments convincing at all. I would put him into the same camp as Deepak Chopra because they make a similar argument that essentially boils down to this:

1. Consciousness is mysterious.

2. Quantum mechanics is mysterious.

3. Therefore, consciousness and quantum mechanics are the same thing.

The key point that he misses is that consciousness (like life, intelligence, animal, and many other biological concepts) is not a binary property -- it is a spectrum. Some things are more conscious than others.

At 7:51 PM, Blogger Lincoln Yu said...

I think they are, at least for now. And 3 should be that many phenomena indicate that they might have something (and possibly a lot) to do with each other.
I guess with respect to the spectrum and non-binary characteristics you are talking about, I don't think he missed it but he expressed it in a different way, which may be much more profound and comprehensive in physical nature.
In general I found Penrose's study though not 100% convincing, but most part of it makes sense, and as far as I know I can't see any more convincing theories than that to me.
One of the only few things I think need more improvement is perhaps the way he uses godel theorm. i guess in his book he was trying to elaborate it for ordianry readers so it occupies substantial chunk of the book which is not worthy of the what it wants achieve. but it's just something that requires more effort. another thing which is more important to me maybe is the lack of critical analysis and discussion of moden AI techniques and technologies (digital or not, i guess mostly digital). He definitely is more than aware of the continuous requirements physics and biology as opposed to discrete ones and devoted quite a bit of his book discussing that as well as how it may correlate with digital computers which is what we mainly have nowadays. and I can fully understand it's no easy job and he never says that's a settled problem.
And I think even he would admit there's a lot of speculations in his work, it's not a complete theory and has lots of flaws and I guess there won't be in the foreseeable future. but his work esp some recent updates (mostly what he collects from the advancements in the areas he thinks relevant) do contain very inspiring views, which specifically guide the research to the underlying physical structures to find out more, and I do appreciate the comprehensiveness and depth of his study.

At 10:30 PM, Blogger Abednego said...

It's not that his theory is incomplete; it's that his theory is not scientific. He offers no testable predictions and no way to validate or disprove his theory. And given that he is a scientist, this is rather inexcusable. For this reason, I put him into the same category as Dinesh D'Souza -- a lot of fluffy-sounding talk, but no real substance.

At 5:34 PM, Blogger RumpleStiltsken said...

Did you read Penrose's papers supporting his thesis? Or did you just think it is so simple that one must be able to teach it orally during a short meeting? Penrose's argument is that electronics cannot mirror human intelligence. He is not saying that machines cannot be intelligent in their own way; they just cannot be intelligent in the way humans are.

At 5:39 PM, Blogger Abednego said...

Does he define the difference between human intelligence and machine intelligence in his papers? A definition of that difference didn't seem like it would be too much to ask for in a one-hour talk. If you're going to claim that machines can't ever do X, then you better at least make an effort to define what X is, instead of appealing to people's inconsistent intuitions about X.


Post a Comment

Links to this post:

Create a Link

<< Home