For FARG's sake. (Fluid Concepts and Creative Analogies - Douglas Hofstadter)

Member Name: mpeh

Product:	Fluid Concepts and Creative Analogies - Douglas Hofstadter
Date:	20/07/02 (129 review reads)
Rating:

Advantages: Erudite, Witty, Enlightening

Disadvantages: Slightly overbearing, A few repetitions

Douglas Hofstadter: Fluid Concepts and Creative Analogies; Computer Models of the Fundamental Mechanisms of Thought

As all thirty of you who read my review of Godel Escher and Bach an Eternal Golden Braid probably know I'm a Hofstadter fan. Although that other book was recognised by the literary community as Hofstadter's masterpiece, he won some major prizes, he has contributed some major work to the modern research community in what used to be called A.I., Artificial Intelligence. In the preface of this book Hofstadter writes: 'In the early 1980's, however, that term (AI) as words are wont to do, gradually started changing connotations, and began to exude the flavor (his spelling not mine) of commercial applications and expert systems, as opposed to basic scientific research about the nature of thinking and being conscious. ... As a result I came to fell much less comfortable writing or saying "AI".' As a result he chooses to use a term that, at that time, was coming into use; Cognitive research. This is a much better way of describing what this book is about, as is the subtitle; Computer models of the fundamental mechanisms of thought.

This book is a collection of research papers published by Douglas Hofstadter with a selection of other people over the ten years from 1985 to 1995. That selection of other people tend to be the research members of FARG, the Fluid Analogies Research Group, either his research students or other members of the group over the years. There is one chapter written by someone not Hofstadter but it fully deserves its place here. The layout of the book is this; each chapter has an introduction written just before publication by Hofstadter and it is these that pull the book together and make it one product as opposed to a random collection of academic papers with one theme. It seems to me that this is a relatively easy way to write a book, after the event that is. Each paper has obviously taken
a lot of time and effort from the people credited as authors and the way in which they are written really takes you through the process of discovery and makes you see the way in which the author's mind was working and how they made the decisions they did. Once these papers have been written, obviously without such a book as this in mind, to put them together with a series of short prefaces to each which tie the book together is a simple exercise and an effort light way to produce a book. This doesn't devalue the book for the reader, in some senses the fact that these are unadulterated (aside from the odd note) academic papers is what makes this book so satisfying. I certainly find when reading 'popular science' books that what often disappoints and annoys me is the feeling that they have been diluted or even dumbed down so that the lay reader, which I imagine the author would count myself as, can cope easily with the content. It's the reason I don't like New Scientist, it tells me one interesting thing, generally gives the abstract to a paper, tells the reader what is being researched and then stops. It just whets my appetite and then leaves off. Too many modern science books do the same. Now I want to be careful here and make it clear that I don't think the majority of modern science books fall into this category, there are plenty I've read that take me far enough into the topic that I feel under educated in the subject area, not clued up enough to keep up with the ideas that the author is trying to communicate to me. So basically what I'm trying to say here is that although the book itself has been put together with relatively little effort the fact that the constituent papers are genuine and unedited and have had lots of effort poured into their creation makes the whole package suitably serious.

The difference between the research FARG were doing and that being done by the rest of the AI research community (as it is p
resented in the book) is in a matter of approach. FARG believed that computers are seriously limited, to be useful in researching the way the human brain truly works they have to be used to model a tiny portion of what our minds do. One of the chapters in the book is an essay, published in a journal at the time it was written, about the comparative claims of FARG compared to what else was being researched in AI at the time. For example one of the research efforts of Hofstadter was to try and make a sophisticated computer program solve a simple anagram in the same way a practiced human being does whereas a pair of contemporary researchers were claiming that they had developed a computer program that could draw analogies between such complex systems and the solar system and Rutherford's model of the atom. Without actually exploring what the chapter itself says in too much detail the crux of the problem here is working out what the computer system actually 'knows' for itself, juggling strings around is worthless as a model of intelligence if the computer has not 'realized' the meaning of the strings, on some level at least, for itself. Obviously the authors of this particular chapter were biased towards their own interests, who isn't? but the points that they raise appear substantial and very valid even under the harshest criticism.

The book has ten chapters, as I said before each is a paper published by Hofstadter and/or his associates as part of their research into the way the human mind tackles relatively simple, or at least more easily modelled, problems. I intend to give a brief outline of each chapter with exciting examples where I feel like inserting them. In an attempt not to give too much away I will curtail many of these.

The first chapter tells of a program designed to pull the essence of a pattern from a sequence of numbers. Mathematicians, and generally anyone with a familiarity for sequences or numbers, can recog
nise the 'rule' behind such sequences as
1 4 9 16 25 36... 1 3 5 7 9 11 13... and even 1 2 1 1 2 1 1 1 2 1 1 1 1 2 1 1 1 1 1 2 ... and extend them indefinitely with only arithmetic calculation slowing them down.
With a little more familiarity or arithmetic effort sequences such as
1 1 2 3 5 8 13 21 44 65... and 0 1 4 27 256... can be recognised and 'decoded' and extended.

To get a computer to recognise these sequences without having a library of known sequences is relatively simple. You program in a series of methods to be followed one after another which basically come down to measuring mathematically the differences between consecutive terms and then playing with these. This doesn't represent the way a person deals with a similar problem. The paper is about the way Hofstadter realised the way people tackle extending sequences and then the way he, and various of his research students and colleagues, coded a computer to tackle the problem in the same way.

The second chapter deals with a program designed to represent the way people who are practised in such things solve anagrams. Apparently, I am not an anagram solving person, people who are used to doing anagrams find chunks of prospective words being presented to their conscious by their subconscious. This issue, interesting as it is, is not what is explored in the chapter. Seeing as this paper was written independently of the first it does cover some very similar ground. This is forgivable for two reasons, it is understandable and introduces the new chapter in an easily approachable way and it is helpful to have the basic ideas iterated as there are some difficult concepts involved here. The way the programs are put together is very interesting and there is certainly enough variation and extension of the ideas already presented that it holds your interest. I would say that this is at a minimum, it much more than held my interest but, as I often find myself doing wi
th something I really like, I am being as critical as I can be here on Dooyoo in order not just to say it's great because because because...

The third chapter is the one not written, or even co-authored, by Hofstadter himself. As with all the others it tells the story of the creation of a computer program to model the way people deal with simple problems. In France there is a game show called 'Le Compte Est Bon' which is very like the number round on countdown in this country. For those uninitiated, or genuinely cool enough never to have encountered Richard Whitely and Carol Vorderman on channel four in the early evening I will here give a description of the game:
The contestants are given four numbers randomly selected from between 1 and 10 and two randomly selected from the set {25, 50, 75} with which, using the four basic arithmetic operations addition, subtraction, multiplication and division, they are required to make their target a three figure number also randomly selected. For example make 437 from 75 25 2 1 8 6

(75-2) * 6 - 1

The French game has a few variations, firstly only five numbers are given; all between 1 and 25 and the contestants may not use division as an operation, the target is randomly chosen from 1 to 150. Again to program a computer to go through all possible combinations with the numbers given would be a simple, and pointless, task. For a more complex problem designing a program which solves it as quickly as possible is a rewarding task in itself but doesn't further our understanding of the way people approach these type of problems. This program takes ideas from both of the previous two and again the chapter, and it's introduction, build upon that which comes before it.

The fourth chapter is the essay about the comparative value of different approaches to AI research which I mentioned before.

The fifth chapter deals with an extension of the program discussed in the firs
t chapter for extending sequences and was designed to solve problems such as:
I change 'abc' into 'abd' do the same thing to 'jkl'. This is a really subtle program seeing as different people give different answers to a lot of these types of problems. The one I gave above had people answering 'jkm', 'jkd' and 'jkl' although this last one is uncommon. Even more variety of answers are given if the problem is: I change 'aabc' into 'aabd' do the same thing to 'jkll' with people finding they relate the double a to the double l and hence give answers such as 'jkmm' instead of 'jklm' which results from simple application of 'change the rightmost letter into it's successor. One of the most popular answers when people are asked to reconsider, or think about the problem for a long time is 'ikll'. The reason people give different answers is firstly because they perceive the initial sequence as being tied together by different rules and then the change itself as being different rules in action. For instance people who answer the problem 'aabc -> aabd as jkll -> ??' with 'jkmm' see the last letter in the first sequence being changed into it's successor (c into d) and then transfer that rule to the second sequence as 'change the last letter (alphabetically) to it's successor' which results in both the double l's being changed into m's; 'jkmm'. Those people who answer 'ikll' are doing so for very subtle reasons. They attach more importance to the doubled letters in both sequences and in some sense equate the aa with the ll, this results in them seeing the change in the first sequence as; 'change the letter opposite the double letter into it's successor'. This idea is reinforced by the alphabetical relations between the letters in the two sequences and would result in the c to d translation in the first sequen
ce being reflected as j to k in the second sequence and would result in the answer being given as 'kkll'. This seems wrong because there is no pair of doubles in the first sequence's image 'aabd'. People get around this, usually subconsciously, by changing the 'replace the letter opposite the double letter into it's successor' by 'replace the letter opposite the double letter by it's predecessor if left of the double and successor if to the right of the double'. This switch from successor to predecessor is made as part of the 'opposite' concept involved with the double letters. This results in the answer I gave before; 'ikll', a more symmetric answer matching 'aabd'. The people who give this answer have morphed the original 'rule': 'change the rightmost letter into it's successor' into 'change the leftmost letter into it's predecessor' via the idea of 'opposite'.

The program discussed in chapter five, called Copycat for the fairly obvious reasons, is in some ways the culmination of the ideas that were responsible for the programs discussed in the first four chapters and as such is obviously a major way post in the research conducted by FARG. As such the next two chapters are devoted to analysis of Copycat, chapter six is a paper comparing Copycat to other recent (when the paper was written) work in similar areas and chapter seven is the after word Hofstadter wrote to a book published by his co author of the copycat program Melanie Mitchell. As such it examines the future extensions of copycat and what use the project was.

The eight chapter is about a program that makes more visual, 'real world' analogies than the sequences in copycat. A program called tabletop was developed to extend the idea of making a computer draw analogies between two situations or within a situation. The scenario is two people, Henry and Eliza, sat opposite eac
h other across a table with different things in front of them, cutlery, plates and cups each with food or drink in them or not. The 'problem' which the computer has to solve is 'Henry points at object A, what would be the analogy for Eliza to point at?'. So if Henry has a plate with a knife and fork set in front of him and Eliza also has a plate with a knife and fork set in front of her and Henry were to point at Eliza's fork then the analogy would be for Eliza to point at Henry's fork (assuming both forks were on left hand sides of plates and both knives on right hand sides). Although this may seem a relatively simple problem the interest comes in trying to make the computer program follow the same intuitive leaps and concept building processes that a human being makes when faced with the same situation. The following chapter analyses the way in which the program Tabletop emerges from the probability based code.

The final chapter of the book is about a text recognition program called Letter spirit. The idea of this program was to see how far fonts or type faces could be altered and still recognised as individual letters by the computer. As with all the rest of this book the idea was to make the computer follow similar pathways to a human and for the realms of acceptable and unacceptable changes to lettering to be the same for the computer as it is for real people. This chapter fits well at the end of the book by being both a conclusion and culmination of the ideas presented throughout and also by not being an absolute ending, as with all real research this isn't the final goal and things don't stop and this really gives you the impression that 'here's where we got to and now here's where we might go:...'.

Throughout the book conclusions referring back to the way in which the human mind actually works are included. FARG was not interested in the neurological makeup of the brain or the way concepts are
dealt with at a physical or chemical level but believe that we can better understand thinking by studying the basic cognitive processes, the bottom layer of information processing at an information only level. Because of the way the book is put together these can be a little disjointed and hence harder to follow than they should be. It would be useful to have them collected together near the end as a coup de grace, a final curtain on Hofstadter's arguments.

The joy of this book is that these are academic papers as they should be written for journals; lucid, witty and firmly based in reality. You have to have your wits about you to read them but the topics they cover and the way that the new advances are presented makes them accessible to everyone. There is none of the largely unnecessary technical detail and so you don't get bogged down in the details. Now don’t misunderstand me here, I totally understand the need for truly technical research and the little details need filling in behind any kind of discovery but when you initial idea is being presented to the public, be they readers of books such as this or a particular scientific research community, you need them to get the overall message and not the working steps. After your research has been accepted, the discoveries understood and contextualised then people will need to come and look at the tables of results, the lists of tests and retests and all the technical details. Further in the future when your results are long accepted and almost required knowledge for study in the field then students will be required to study your reams of results to see where your conclusions and ideas were drawn from but at the beginning, right at the edge of discovery the excitement should be allowed, even required, to shine through.

On a more critical note it is in the same area that the book could be improved. By not editing the papers he is presenting at all I feel Hofstadter has missed an oppor
tunity to embellish upon the explanations given in the text. There are times when a thoroughly worked example or annotated run through of one of the computer's runs would have been very helpful and it surely wouldn't have been much effort to put them in. Also the prefaces he writes for each chapter, although being a good idea could be better executed. They provide an historical (in the recent sense) background of who wrote the paper and the scientific/ political climate at the time as well as introducing the concepts being explored but they could be brought into playing a major role by simply having an after word to each chapter, a review of what was discovered and what it meant then and means now. Even just a précis of the major points made in the paper.

This book is a really good read, you are very unlikely to sit down and read it from cover to cover and it is no more an up to date representation of alternative research in AI. Because Hofstadter and his associates were, and presumably are, researching cognitive science in a very different way to the rest of the AI community they inevitably have differing views and, equally inevitably, believe their approach to be more valuable or realistic or just worthy. As such some of the chapters in this book, especially the critique of AI research at the time and the comparison of Copycat with recent work are biased. This would be true of any similar work but it can be a bit jarring to be reading something which is so lucid and then only a few pages later realise that the point of view being put across may not be the only valid one. Hofstadter occasionally seems taken with his own brilliance, which, although undeniable, I didn't like having rubbed in my face. He happily gives credit to all the other people who worked with him and around him during the various projects but most of the prefaces include a story about how the initial concept, such as playing with specific numerical sequences or seeing how fa
r a typeface can be stretched and yet still recognisable, came about through his being interested when he was a child or teenager. I am sure that these are not included to annoy the reader or to make them feel less clever than the author but the repetition of this theme can become annoying.

If you are an Hofstadter fan then you will find this another delight. Written with his usual flair for puns, wordplay and 'characteristic wit and brio' it is easy reading which as usual educates as well as enthrals. Some of the conclusions drawn may seem a little stretched or odd but remember this is eight or nine years old now and also that we don't have access to the full research which led to these conclusions. I feel Hofstadter reveals much more of himself in this book than in others I have read and yet I am torn as to whether this is a good thing or not. Certainly a passionless book without any of the author in would have been less preferable but maybe a slightly more objective view of the papers would have sat better with me.

Summary: