Home
Sometimes I think intelligence is nothing but pattern recognition. Someday in the near future a computer scientist will write code that rapidly compares and stores complex patterns. To populate the computer model with data, the programmer will let his software read the entire Internet, or as much as it can, and look for patterns. After a few days of chewing through content on the Internet, the software will appear to understand everything about humans, from our language to our history. The reality of this apparent "understanding" would be nothing more than pattern recognition.

Consider a simple example: A computer could learn by pattern recognition that humans raised in specific parts of the world usually say "bless you" to a person within earshot who has just sneezed. Then overlay the pattern that people usually whisper in a library and your artificial intelligence knows it should whisper "bless you" when someone in a library sneezes. In a more complicated scenario there might be hundreds of patterns intersecting, but a computer could easily contrast and compare them.

I could be wrong, but I suspect that artificial intelligence will grow out of one page of clever pattern recognition code combined with exposure to every pattern revealed on the Internet. It would take about a week to turn the pattern recognition code into what appears to us a sentient being with super intelligence. Futurists call that day the singularity.

As with most of my posts, some of you will tell me about all of the fiction books that say the same thing but said it first. I haven't read any of those books. Nor do I know anything about actual AI research, so perhaps it's obvious that pattern recognition is the key, and the real problem is that it's a hard nut to crack.

I came to my hypothesis that intelligence is just pattern recognition because people who are not terribly bright have trouble understanding analogies. And analogies are just patterns.

Logic isn't a big part of human intelligence. Put three humans in a room with a problem and each will have a different idea of the logical solution. Humans are rationalizers, not logical beings. Computers don't need logic to act human because humans don't have enough logic that anyone would notice some was missing.

Experience is little more than having more patterns to draw from. New situations are never identical to old ones, but they might follow a general pattern. For example, when I was in my twenties and someone said they would call me back I assumed it was true. Now I look at the entire situation, and all the patterns involved, and half the time I correctly tell myself I'll never hear from that person.

My two questions for the day:
  1. Is intelligence much more than pattern recognition?
  2. If intelligence is mostly pattern recognition, how likely is it that someone will write code that scours the Internet for patterns and uses that as a base to create super intelligence?
 

 
Rank Up Rank Down Votes:  +25
  • Print
  • Share

Comments

Sort By:
+1 Rank Up Rank Down
Mar 1, 2014
PS -- That grawlix in my posting was supposed to be the word 'ent_ities'.
 
 
+3 Rank Up Rank Down
Mar 1, 2014
A crucial difference between human intelligence and pattern recognition is that human intelligence is also informed by a POINT OF VIEW. This requires several things:

1) Conscious self-awareness
2) Motivations
3) A physical context for one's consciousness -- in other words, being located in a particular space and having a physical body.
4) A social context for one's consciousness -- in other words, the knowledge that other beings exist, and that they also have minds, emotions etc. which may differ from one's own.
5) An explicit organizing framework for understanding the world which heavily involves, but is not limited to, the use of language.

Of course, there are many simple problems that CAN be analysed and solved by logic and/or pattern recognition alone. But an artificial intelligence can never truly UNDERSTAND the data it is analysing, because it is incapable of relating any of it to its own needs and feelings (since it is incapable of EXPERIENCING those things). The best it could do would be to generate a facsimile of human understanding, but it would be unable to make the same kinds of creative or intuitive leaps that characterize the highest levels of human intelligence.

(This same lack of self-referential and social understanding is also key to the impossibility of perfecting language translation software except in very restricted contexts, such as translating parts lists; I do not believe that a non-sentient algorithm can EVER be as good as an experienced and knowledgeable human being at translating ambiguous source texts, even granted that human translators also make mistakes. Indeed, that would be a very good task to set as a Turing test.)

Finally - if human intelligence was truly reducible to mere pattern recognition, our lives would be arid indeed, because we would have no way of making them MEANINGFUL to ourselves; we would be little more than a bunch of databases attached to data-gathering sensors. Without the capacity to experience meaning, such mere pattern-matching intelligence can only be useful if it serves the needs of !$%*!$%* that DO have that capacity.

Of course, human thinking is nevertheless highly flawed; the incremental nature of our evolution means that sophisticated analytical thinking is easily short-circuited, overtaken or skewed by perceptual anomalies and by non-analytic instincts and brain functions.
 
 
Mar 1, 2014
I think that similar method is used when developing machine translation system, surprisingly (at least for my, quite complicated language), it doesn't work - in my opinion, there were improvements in the beginning (some 5 years ago), but now the results in recent time is less and less satisfactory.
So, how could be created "genius robot", if the scientists (for now) failed in the sub-task of creating reliable machine translation?
 
 
Feb 28, 2014
It goes beyond pattern recognition to weighted pattern recognition and pattern forming and just understanding the world.

People say "Hi", "Hello", "How are you" and many other things when greeting. An AI that recited all of them when it met someone would seem crazy. And an AI that figured out how often people said each one and randomly said one would seem normal at first but then seem random. Imagine if an AI decided that 1/4 of the world spoke Chinese, so 1/4 of the time it spoke in Chinese.

When you think it is just pattern recognition that is because you automatically assume all of the other intricacies and factors that an AI would have to know.

To use your example the AI would have to know: what is a library, what is a sneeze, how loud is a whisper, how far away can someone be and you say "bless you", how far away is the person who sneezed.

[I think enough information and enough patterns solves all of the problems you mention. -- Scott]
 
 
0 Rank Up Rank Down
Feb 28, 2014
Chess is a good counter-example. Or at least an example of where pattern recognition is too hard.

Computers play chess by brute force. They look at as many possible future positions as possible - and the further you go, these of course multiply exponentially.

A computer can't play like a person. They can't just "see" things. In fact a computer playing with a restricted CPU (or just an old computer) can be beaten by confusing it, luring it into positions where long-term, subtle traps can be laid. Even here, the computer can look at millions of future positions, a vast advantage as that's millions more than any person can, but the person has the tools to discount 99% of choices without needing to evaluate them.

Now, either people have some logic/higher intelligence/something else at work here, that machines do not, or it is that the patterns are too complex and too inter-related for the machines to copy.

[Chess actually supports my point. Humans use patterns and today's computers use brute force to calculate all paths. I'm suggesting that computers will someday be built that can recognize patterns as well as a chess master, thus not needing to calculate every path. -- Scott]

 
 
Feb 28, 2014
Sorry. Typo in the first sentence. It's supposed to read, "Lots of MEAT (robots?) here.
 
 
Feb 28, 2014
lots of mean (robots?) here.

I have worked in AI, so I can speak somewhat intelligently about it (pun intended).

One of the things humans do well and computers do poorly is pattern recognition. For this reason (among others), software that can understand natural language commands is very difficult to create, and it doesn't work very well.

Think of how you can see many different cursive documents written in many different scripts and still be able to read them all. A computer has a tough time with that. Along the same lines, you can listen to a number of people speak English, each with a different accent, and still understand them. That's very difficult for a computer.

Have you ever used the "Dragon" speech recognition software? That's one of the best there is, and it's not all that good. It first must 'learn' your speech patterns (interestingly, one of the reading examples is a few paragraphs from one of Scott's books!), and then it requires a lot of work to get it to do a half-baked job of translation. If you try to run a digital file through it instead, the results border on the hilarious. The later versions take an enormous amount of CPU power, to where your system runs very slowly.

Neural networks show a lot of promise in doing things that humans are good at. Give them a problem like the 'Traveling Salesman,' where the most efficient route between a certain number of stops must be computed, and they can give you a very good answer in a short amount of time. A standard computer can give you the very best answer, but it takes orders of magnitude longer.

Another problem computers have is with the imprecision of (particularly spoken) language. Consider the phrase, "Time flies." Think of how many different meanings those two words can have.

A lot of our understanding of spoken (and written) language comes from context. We know what a word or phrase intends to convey because we remember what came before it and can therefore put it in perspective. Computers have a very difficult time doing that.

Think of the sentence, "She found a rose on her car window." Now think how different the meaning, and the emotional impact, of that sentence is in different contexts. If it's a love story, where the rose has a special meaning to the woman, you get all warm and fuzzy inside. If it's a murder mystery where the killer leaves a rose on the car of each woman he's about to kill, you shudder inside. A computer would be clueless.

So overall, Scott, I'd have to disagree. There's a lot more to intelligence than a computer's ability to recognize patterns and then get a plethora of information off the Internet. I recall a very large California bank who tried something similar: this bank made a lot of its money through currency trading. The problem they were trying to solve through AI systems was as follows:

They had trading desks in three areas, to take advantage of 24-hour trading capability: Tokyo, San Francisco and London. Traders could only read at human speeds, and so different traders may pick up on different indicators of whether they should buy or sell a particular currency. If a trader in Japan got a 'sell' indication on Yen, while a trader in London got a 'buy' indication, then the trades would cancel each other out, and actually cost the bank money from the transaction fees and overhead.

So the bank wanted to build an expert system to do just one thing: read indicators at computer speeds and look for buy or sell patterns. When it detected one, it was to flash an informational message recommending against the counter trade. In other words, if the computer ferreted out a 'buy' indicator for Yen, it would flash a message to other traders in all countries not to 'sell' Yen until the indicators could be evaluated.

Sounds easy, right? Wrong. It didn't work.

I'll leave you with one apocryphal AI story from many years ago. During the Cold War, a hot line was established between Washington DC and Moscow to prevent an inadvertent launch of nuclear missiles by allowing rapid, direct communication between the two heads of state.

There were two problems: first, the odds were that the US President didn't speak Russian, and vice versa. The second was that the topics discussed would be of such a highly classified nature that it would be problematic to let translators listen in.

So the two governments embarked on a years-long project to build a machine translator from Russian to English and English to Russian. Finally, the machine translator was ready to test.

Fully understanding the problems associated with language, the team picked two challenging phrases. The first was translated from Russian to English and then back to Russian. The second was translated the other way.

The first phrase was, "The spirit is willing but the flesh is weak."

The second phrase was, "Out of sight, out of mind."

The first phrase, after translation, came back, "The Vodka is good but the meat is rotten."

The second phrase came back, "Invisible insanity."

The project was scrapped.
 
 
+3 Rank Up Rank Down
Feb 28, 2014
Yea, you pretty much nail it: pattern recognition is fundamental for a working AI and it is a very hard nut to crack. The problem, to be brief, is: you need to form abstraction, you gave behavioral abstraction with "bless you" and the "whispering in a library", those tend to break with other problem. Thus you need multiple forms of abstraction. Then you need to classify what you find on the Internet, and chose, when acting, which form of abstraction to use. Nothing impossible to do, but it is really complex. Human are like a 1 or 100 billion core cpu clocked at 5 or 10Hz (not Kilo, or Mega or Giga, just simple Hz). Right now, it is a nightmare to create a program that use more that 4 thread safely. Well, there is graphics, but that just one problem that scale very well.

1. Yes, it is a definition for intelligence. Do note that research indicate that in the last century, the IQ rose quite sharply and people are now better at pattern recognition. It is all about pattern and abstraction you can draw from those pattern and then reusing them elsewhere.

2. Very unlikely, the closest thing we have to this is Google and their search engine is quite far from being an AI. It is surprisingly useful, yet I would not qualify it as very clever. However you are probably right with the idea that the first pattern database would be create from Internet data.

[I think a good pattern recognition program could start without a sense of language and learn it by pattern recognition alone. Our minds tend to think within the confines of language, so the computer would have that structure in place. After that I think pattern recognition alone would solve every "if then" situation at least as well as humans do it. I think patterns would even confer some basic morality on the program by example. -- Scott]
 
 
Feb 28, 2014
Our brains are wired for pattern recognition. Most of the hard wiring comes from the first few years of life where the neuron patterns of speech, facial recognition, walking, talking etc are laid down. There follows the educational years where the patterns are reinforced and acceptance in your peer group happens by behavioural patterns that are the norm within your group. Hence we get 'my side' and 'your side' spats developing as 'we are right and you are wrong' theories develop and alignment to certain sports venues becomes part of the life of the individual. The pattern of strength in numbers develops as we seek identification with neighbourhood/gang/district/nation. Seek for patterns in the web and you get a mess of conflicting (and sometimes very nasty) views/punishments for not belonging to THIS or THAT group. How do you choose the 'right' data?
 
 
0 Rank Up Rank Down
Feb 28, 2014
For a vaguely related idea, see here:

https://twitter.com/cmunell

It's a bot that tweets facts that it infers from crawling the internet at large.

That's not the same thing as using these facts to create patterns as a foundation for AI, but it could be the start of one. I see a lot of AI projects heading in this same direction, and I would be surprised if IBM's Watson didn't use concepts that are similar to your ideas.
 
 
0 Rank Up Rank Down
Feb 28, 2014
I think you are mostly right about intelligence and pattern recognition, but not about Internet.
Drawing data from internet without a fact-checking mechanism and a sound theory of reality
will produce a dumbass intelligence believing in reptilians, chemical trails,
mass destruction weapon threates and other such hoaxes.
Information without education is useless.

You could be anyway right, if human intelligence should decrease down to an Internet-driven robotic level.

Of course, you could obtain a sort of Internet-driven intelligence,
well adapted to survive in Internet,whatever this means (it might mean something, I guess).
But tranferring all this to the physical world is yet another step.
 
 
Feb 28, 2014
I like this post. Fun to think about.

It seems like humans are great at recognizing patterns of intelligent behavior in most other creatures.

When it comes to us, there comes a step where we lose the ability to rationally explain what is happening. For example, a human performing normally was using intelligence, logic, and past experience to complete test questions. Someone performing remarkably, however, is somehow using a "leap of faith" to achieve their remarkable performance.

It seems the "any sufficiently advanced technology will be indistinguishable from magic" applies to our brains as well: we can't explain them- no surprise. My guess is someone who could would seem full of nonsense or crazy to the rest of us.
 
 
Feb 28, 2014
Sure - patterns are all you need to create the world's smartest paperweight.

You need motivation. Hunger, fear, love, lust, etc., drive humans forward.

I'm not saying the Id, Ego, and Super-Ego can't be programmed - I'm just saying that without them your robot won't have a reason to get out of bed in the morning.
 
 
Feb 28, 2014
Chess is really just pattern recognition. and compiling future outcomes based on present scenarios. When a human becomes a master at chess, we typically believe this person to be extremely intelligent.

Here's another quirk about computer-based super-intelligence...since there is so much data about the stock market, how come we haven't been able to develop a program that recognizes patterns and predicts stock performance with a high degree of reliability?
 
 
+2 Rank Up Rank Down
Feb 28, 2014
That's certainly true in the field of algorithms.

I had a coding interview recently, and one of the questions was how to find a missing element from an unsorted list of size N comprised of unique numbers within the range 1 to N 1. In other words, which number would have to be added for the list to contain all numbers between 1 and N 1? I saw it as a variant of a problem called "find the midpoint", and gave a provably correct answer. What the interviewer wanted was something more clever, using Euler's sum of an arithmetic sequence to figure out which element was missing. If I'd had better pattern recognition, I might've caught it. Afterward, I coded it both ways out of curiosity, and found that my answer and the correct answer both ran in linear time, but the correct answer ran almost exactly 20 times faster for every possible size of input.

Still waiting to find out if they're interested in hiring me. My pattern-recognition says it's possible but not likely.
 
 
Feb 28, 2014
I can only hope and pray that these computers are NOT allowed access to Youtube comments.
 
 
+2 Rank Up Rank Down
Feb 28, 2014
Google's spell check, "did you mean" and language translation all work this way. Google scraped enough web content to see how people spell and mis-spell words (in every language for which spelling is meaningful). It also finds existing translations of works into other languages to figure out how to machine translate things.

If you did want to read a sci-fi book with this theme - then I'd highly recommend Daniel Suarez "Daemon"
 
 
Feb 28, 2014
I was going to try and quote what IQ measures, but I can't seem to find the list today.

So, forgive me for being inaccurate, but IIRC... IQ measures what is considered to be the set of cognitive skills that most affect how quickly a person can learn something. These are: memory, perception, pattern-recognition, abstraction, and logic.

I think memory is something we can assume for a machine. Logic is a gimme, too. Perception might be a challenge, because it means a lot of different things. Pattern-recognition is an even more vague concept. Abstraction is something we can barely teach people.

That's not even necessarily to say that these are a comprehensive set of relevant cognitive skills. There might be dozens or hundreds of others that are necessary at certain thresholds, or aren't generally recognized, or are less useful for the purposes of a test like this.

I believe we'll make important breakthroughs and develop increasingly useful machines with surprising capabilities in the immediate future, but human-like intelligence seems like a very distant goal. I am not remotely convinced that we know how to describe it, let alone construct a system for it and simulate it.
 
 
+16 Rank Up Rank Down
Feb 28, 2014
Wouldn't a real genius scouring the internet for information (aka - patterns) be even slightly interested in prior research (aka - books already written by researchers who may of had more than a passing thought on the subject)?

Scott, why so stubborn (and maybe proud?) about not reading books on subjects which you're clearly interested? Not trying to argue, just often wonder at this whilst reading one of you posts.

Wouldn't "super intelligence" at least start with what is already known and build upon it?

And hers is three more question marks, in case I forgot one -> ???
 
 
0 Rank Up Rank Down
Feb 28, 2014
Maybe singularity will be achieved when our use of the internet has dumbed us down sufficiently.
 
 
 
Get the new Dilbert app!
Old Dilbert Blog