Fluid Concepts

pages 195 to 204

2009-10-27T02:19:00.000-07:00

After the fact, I can understand the "logic" behind the Italien breakup of "in fact" into in effetti, di fatto, infatti, in realtà, anzi, però, tant'è vero che, per esempio, a dire il vero, and so forth, but I never would have thought of it myself.

(page 200)

Obviously no exact number of concepts exist for any occurrence whatsoever.
The quality and quantity of concepts depend strongly on one's culture and environment.

In an environment with very frequent snowfall one may have more words for the different kinds of snow than in other places; a society living in an environment with many different sorts of sand most likely will come up with a respectable amount of words describing these different types.

Also worthwile pointing out is Hofstadter's idea of conceptual slippage:
For example, the Italian question "Lei ha fratelli?" actually means "Do you have brothers?" (page 199). But what if one has two sisters instead? Will one simply answer "No."?
Of course not; In English one would answer "No, but I have two sisters.".
And in Italian one even answers "Sì, due sorelle" - meaning "Yes, two sisters.".

Why is this? The reason is that many statements and questions (like here the "Lei ha fratelli?") are merely to be understood metaphorically.
Every real-world occurrence is more or less "fuzzy". There is (practically) nothing that is 100% this way or 100% that way. Therefore, our concepts must be "fuzzy" as well.

An artificial system that should even remotely be able to interact with its environment in a human-like manner must account for this fuzziness in order to be successful.
This fuzziness must also include the ability of constantly being able update one's concepts, if necessary.

pages 169 to 193

2009-10-26T22:42:00.000-07:00

Now that I am writing, it is essential that I conceive my paper as a surface for inscription.... But if I wished to light a fire, and no other materials were by, the essential way of conceiving the paper would be as a combustible material....

(page 174)

This quote shows very clearly that our active mental representations at any given time are highly context-dependent.

We can - and we do - constantly perceive similar objects in completely different ways:
a liquid may be drinkable in one situation and used to distinguish a fire in another;
a glass may be something to drink out of in one situation and used as a weapon in another.

It is highly unlikely that we constantly are "loading" all concepts into our active mental representations of present occurences. Instead it seems plausible that we only use those concepts that are relevant to a current situation.

Perception and cognition are constantly influencing each other.
If we are thirsty, we will search for drinkable fluids. If we have to put out a fire, we will search for non-inflammable fluids. If we want a container to drink out of, we will usually search for different objects than if we need something to mix glue in.

Many A.I. research projects provide pre-coded representations to their programs.
But due to the above mentioned statements this, most likely, may not be the best approach.

In order to model the human mind, perception and cognition must interact with each other in order to create context-dependent - i.e. flexible - representations.

Trying to create this type of artificial system from scratch would pose immense computational problems because the real world offers an unbelievable amount of information every second throughout various modalities.
Therefore Hofstadter suggests to create microdomains in which artificial systems have a manageable amount of data available and where perception and the creation of representations can interact with each other.

After having read this chapter I definitely agree that this approach sounds highly promising.

pages 155 to 168

2009-10-26T15:14:00.000-07:00

a compound word like "knows-truth-or-falsity", so transparently evocative to readers of English, might as well be, for all the computer could care, "xjs-beuglh?" or "doesn't-give-a-damn-about" or the digit "8" or any other alphanumeric string.

(page 164)

This excerpt shows most obviously that todays' computers simply are not able to actually understand anything at all even in a remotely similar way that any average human is able to.
We can pass arbitrary values to computers which can store them in variables. Then the computer can do various computations on these variables; but no matter how complex these computations may be, there isn't even a spark of actual understanding involved in this process.

The only things today's computers are capable of doing is "symbol manipulation".
When we read the output of computer programs we often are impressed and have the feeling that what we read actually makes a lot of sense. We tend to believe that the program which created this output is very clever.

But we may never forget that however smart the output may appear to us, it has absolutely no meaning at all to the computer program which created it.

pages 139 to 154

2009-10-10T14:45:00.000-07:00

Fighting Codelets

There seems to be no central processing center in our brain but rather a

probabilistically-biased parallel exploration of possibilities.

(page 150).

The principle of collective intelligence is based on the principle that many individuals will nearly always come up with a better solution than any individual on his or her own.
Constant generation, regrouping and chaining together of new ideas has the potential of always finding a satisfying solution.

This is exactly what is going on in our brain:
We are constantly producing "codelets" (small parts of information), regrouping them and chaining them together in various fashions until the solution to a problem is found.
Codelets are constantly competing with each other, trying to come closer to solving a given problem.

For our purposes the only obvious difference between humans interacting with each other and codelets interacting with each other is that codelets are not self-aware. They do not have the "big picture". Rather the interaction amongst them is actually creating the global picture.

The precise strategies of these interactions remain unknown.
But Numbo is definitely a good starting point in order to try to understand more about this emergent phenomenon.

pages 131 to 138

2009-10-10T09:57:00.000-07:00

Human Problem Solving: Knowledge or Randomness?

Some problems can be solved by applying knowledge alone: If we are experts in one field or another, we have no doubt that this or that problem can be solved in a very specific way which are previously aware of.
On the other hand completely unknown situations may make it necessary to generate strategies on a random basis and simply hope for the best.
But on a more general basis it seems quite straight-forward that most problems we have to deal with are neither solved solely by applying previously acquired knowledge nor solely on a random basis.
Almost always we at least have some kind of intuition what may be the best way to deal with whatever situation we may encounter.
Intuitions seem to consist of small knowledge fragments which may or may not be applicable to the current situation. But what is important is that intuitions very often lead is in the right direction. One may call this an "educated guess".
This is exactly what happens when we try to solve problems:
We try to use as many fragments of previously stored knowledge as possible; and whatever may still be missing is added on a random basis.

pages 112 to 126

2009-09-26T12:47:00.000-07:00

Thermodynamics and Thinkodynamics
or
Organized Disorder

One central question of research on cognition is how precisely we must model the human brain in order to model human cognition.
In thermodynamics we don't have to know about the individual atoms and molecules. Rather we apply statistical measures to calculate the macro level of a system.
Hofstadter considers applying this concept to thoughts and calls this principle "thinkodynamics".
What does this mean with respect to cognition?

The traditional holy grail of AI has always been to describe thoughts at their own level without having to resort to describing any biological (i.e., cellular) underpinnings of them.

(page 125)

Future research will have to show if this claim is actually true. So far neither the low nor the high levels of the essence of cognition have truly been found.

However it does seem quite plausible to imagine some kind of "organized disorder" takes place on lower levels of cognition. Perhaps various sub-structures of a cognitive system constantly run in parallel and produce educated guesses, led by statistical analysis. Depending on their actual relevance with respect to any current situation the mental building blocks created in this manner are transferred to higher levels.
Hofstadter's theory is that the higher the level of cognition is the less parallel the processing is.
Processes on the highest level of consciousness may occur fully in serial.
This claim seems absolutely valid. Of course, we can do several things in parallel.
But although depending on the situation each thought may follow another in extremely short succession, it does seem to be the case that each conscious thought actually occurs one after another.

I want to end this entry with the following question:
Could the seemingly organized structure of our top-level cognition be a merely coincidental result of emergent properties of more or less chaotic structures on the lower levels of our cognitive system.

pages 97 to 111

2009-09-24T06:05:00.001-07:00

From this section of the book I most enjoyed the concept of a "terraced scan".
Hostadter uses the example of going into a bookstore and rather than reading each and every book cover to cover before deciding which book one is actually interested in, one tends to sort out so very many books by the blink of an eye. We are able to categorize things at extremely high speeds. Of course we are sometimes wrong but in general this kind of approach facilitates our everyday life quite a lot. In principle only the "interesting" reach higher levels of cognition.
However, Hofstadter pointed out that this sorting out is actually not a binary decision; instead every "thread" returns a value indicating how interesting or not interesting an occurrence is.
Therefore if at first glance, occurrence A seems to be the most interesting but after more thorough research turns out to actually not be, one can anytime return to occurrence B which at first was categorized as less interesting.
This is one of the core principles that Hofstadter is trying to implement into his artificial intelligence research project "Jumbo".
Basically this approach is "merely" dealing with search strategies. I know that various A.I. researches believe that intelligence is "merely" achieved by efficient search strategies.
But I am not so sure about this. I am sure that research on search strategies will be able to significantly enhance artificial systems. But I doubt that it will be sufficient in order to create anything truly intelligent. Emotions and intuition seem to me to be equally important.

pages 87 to 95

2009-09-18T09:13:00.000-07:00

In this section Hofstadter uses the word play with anagrams as an example for learning something about the way human cognition works.
A lot - if not most - of our cognition occurs unconsciously. If we see the character string "4 + 5" somewhere we automatically parse it into the meaning of adding a quantity of five to a quantity of four and calculate the result. Of course this only occurs if we are able to read and have learned during some previous time about the concept of numbers and addition.
But as soon as one has gained this kind of knowledge one is not only able to perform this task automatically; in fact one can't even prevent this automatic analysis of happening.
Some subconscious levels of our cognitive system constantly analyse incoming signals and if these signals match what the area in question is specialized for a message is sent to higher cognitive areas bringing this finding into consciousness.
The same thing happens with anagrams. Hofstadter describes this phenomenon as to "sit back and watch" - because the level of influence we can take on the type of processing occuring during the building of anagrams is not really as high as one might tend to think.
Modern programs tend to use exhaustive search algorithms for all kinds of problems. These programs are not only capable of finding all possible solutions within their designated domain but they can also perform this task at incredible speeds.
However this type of processing is not even remotly similar to the way human cognition works.
Therefore Hostadter writes with respect to these types of programs:

I have little interest in them, aside from genuinely admiring the clever hacks
involved in programming them.

I strongly believe that this claim is very important. Because it becomes ever more obvious that our human mind works in a completely different manner than that.
Hofstadter introduces a concept of different levels of analysis, with the most fine-grained level running in parallel by using a bottom-up approach while specific higher levels running in a top-down manner are only triggered when appropriate.
All this happens totally automatically and intuitively. The structures underlying these automatisms and intuitions are precisely where the emphasis of research must lie.

pages 70 to 86

2009-09-16T16:59:00.000-07:00

Generalization, i.e. inductive reasoning, is undoubtedly one of the key capabilities of human intelligence or intelligent systems in general.
How does generalization work? First of all one must form concepts to begin with.
I like to think of a concept as a collection of ideas governing one entity or another.
Very deliberately I chose the word entity. Because concepts don't even have to be representations of real world occurrences. One must also consider all the kinds of mythological and theological thoughts that can arise.
However I enjoyed reading about Hofstadter's view on concepts very much, interpreting them as spheres with representations of most typical conceptual examples in the core and less typical examples forming the outer layers. Least typical examples which may already count as examples for other concepts form the sphere's blurry edges.
There is undoubtedly quite a lot of generalization going on in the everyday life of each and everyone of us. Just think about the abundance of incomplete sentences uttered in natural language.

Computer scientist: I'm in artificial intelligence because it's a mixture of psychology, philosophy, linguistics and computer science.
Architect: That's the reason I'm in architecture.
(page 76)

And I believe that this that is exactly the reason why artificial systems will never be able to grasp what a human being is trying to actually convey with his or her utterances before it actually has some intelligence of its own.
This that in question does not even refer to any specific linguistic constituent of the utterance of the computer scientist. It rather refers to a concept - the concept of enjoying one's job due to enjoying its underlying intellectual framework.

As far as I know so far nobody has come up with any solid idea on how to model this kind of ability to generalize in an artificial system.
But in order to be able to create something truly intelligent this is definitely one of the starting points we have to look at.

pages 55 to 70

2009-09-15T03:29:00.000-07:00

In this part of the book Hofstadter writes about pattern-finding, analogy-making and the difficulty of finding out which parts of any given information are central and which are not.
Research has shown that there are many, many processes going on in parallel. Hofstadter describes these individual processes as islands (acquired by bottom-up processes, i.e. data-driven) which have to be connected by perceiving relationships (by applying top-down processing, i.e. goal-driven).
I believe that modeling this massive parallel processing is the key to enable an artificial system to act in an intelligent manner.
We will need to find out just what exactly are the smallest information units (i.e. islands) and which rules govern the process of deciding which islands are to be linked and which not.
I also enjoyed Hofstadter's analogy with respect to comparing human cognition to an ant colony very much. It seems dazzling to find out just what exactly which member of an ant colony is doing and which goal each member is trying to achieve. It is even ever more fascinating to find out what the colony as whole is capable of doing by merging the actions of the various members into a larger framework.
This kinds of thoughts shall be our starting point for further research...

pages 35 to 55

2009-09-14T19:22:00.000-07:00

One important aspect addressed in this part of the book is the question what the key to intelligence is; is it knowledge, a knowledge-independent core, a combination of both or perhaps even something different?
Whatever the answer is, I believe that so much is probably true: We won't be able to put our fingers on any single aspect. Given the extremely complex structure of our brain, a simple answer simply doesn't appear conceivable.
However, it seems likely that intelligence emerges from a very sophisticated interconnectedness of various underlying substructures.
My first thought was that knowledge must be one of these substructures. After all it is the case that without knowledge one is lost in time and space - without being able to integrate sensory inputs into meaningful patterns. But then again I can imagine a system to start out completely without any knowledge at all and to emerge into an actually intelligent system merely by analyzing the incoming patterns via a "knowledge-independent core". Immediately after birth/creation of such a system the agent may not be intelligent yet in the way we think of intelligence. But I am willing to call a system intelligent even if it is only potentially intelligent, i.e. if it has the inherent ability to gain intelligence.

Introduction

2009-09-07T20:35:00.000-07:00

Studying Cognitive Science means having the wish to understand how cognition actually works.
One major cornerstone in this endeavor is Douglas Hofstadter. He believes that the core principle of intelligence is the ability to form analogies.
Concepts are not static but rather "fluid". Ideas and theories about one entity or another can therefore also be applied to previously unseen entities.
Trying to apply this principle to realm of artificial intelligence remains widely unsolved:
Computers simply do not "understand" in a strong sense what they are actually computing.
Nowadays it is therefore merely impossible for a computer system to abstract knowledge from one domain and apply it to another.