•Catalase Website Index •James •Techno •Thought •Thought Mini-Review •

Approaches to Mind - A Selective Review

In the following sections I've given my reactions to ideas a number of researchers have put forward on the mind. There is no deliberate distortion in this, but I may well have mis-represented what each of these people is saying, since I pick out what speaks to me. Go to the original sources wherever possible.

Overview

Functional Units Aleksandr Luria recognised three main functional units in the organisation of the brain
Virtual Serial Machine Daniel Dennett examines the differences between our perception of our inner processes and the inner processes themselves.
The Baldwin Effect Evolution has a greater chance of discovering flexible systems than inflexible ones.
Cognition as Compression Gerry Wolff sees acquisition of information as being allied to information compression.
Language Instinct The subtle complexities of language and the inability of apes to acquire it in any real sense, Steven Pinker argues, point to special purpose linguistic mechanisms unique to humans.
Postscript Why multiple points of view are needed.

Overview Functional Units Aleksandr Luria recognised three main functional units in the organisation of the brain Virtual Serial Machine Daniel Dennett examines the differences between our perception of our inner processes and the inner processes themselves. The Baldwin Effect Evolution has a greater chance of discovering flexible systems than inflexible ones. Cognition as Compression Gerry Wolff sees acquisition of information as being allied to information compression. Language Instinct The subtle complexities of language and the inability of apes to acquire it in any real sense, Steven Pinker argues, point to special purpose linguistic mechanisms unique to humans. Postscript Why multiple points of view are needed.

•Top•Down to Dennett•

Functional Units

The Working Brain - Aleksandr Luria ISBN 0-14-08.0654-7 (1973)

The neuropsychologist Aleksandr Luria was particularly interested in the physical location of different cognitive functions in the brain. He studied people with strokes and tumours and collated observations by many different people working in the field.

In 'The Working Brain' numerous case studies are used to illustrate the spatial variation of function within the brain.

According to Luria, the brain has three principal functional units:

For regulating tone or waking
For obtaining processing or storing information
For programming, regulating and verifying mental activity

The first unit controls awakeness from asleep to wide awake. It also controls sensitivity to external stimuli. When we're on edge, we react more to quiet sounds. The first unit initiates the orienting reflex which directs our attention to an unexpected stimulus. Luria points out that it down regulates excessive activity too.

I am unsure whether I have correctly understood the distinction Luria makes between second unit and third unit. By my reading the distinction seems to correspond to subconscious mind (second unit) 'what we do automatically' and conscious mind (third unit) 'what we are most consciously aware of thinking'. Another way I read this is that the third unit writes and modifies 'programs' and the second unit executes them.

Comment
Is Luria's division into functional units a useful one? An interesting piece of research suggests that for facial expressions showing emotion the 'principal component' we react to is not on a happy-sad axis but on an asleep-awake axis. This suggests Luria's observation that the first and 'oldest' functional unit of the brain controls alertness levels is an important one.
For me the distinction between second and third unit is unexciting. As a programmer I am already prone to overpolarise program vs memory, and I would far rather something that pulled me in a different direction, that made me think about the problems of the mind from a different viewpoint.
However, I like the singling out of the first functional unit which, like a searchlight, directs attention and prevents the brain being under or over active. In my own attempts at 'smart' systems (for laying out IC designs) I found I had to explicitly recognise overactivation of subsystems. I did this before reading and understanding Luria. With that background it now seems only natural to me that the brain, being a major consumer of metabolic energy, will have as a fundamental systems some mechanism for regulating use of its own resources.

•Top•Down to Baldwin•

Virtual Serial Machine Consciousness Explained - Daniel C Dennett. ISBN 0-14-012867-0 (1991). A crucial observation is that our perception of our own internal processes is not a reliable guide to how they work. For example, with our visual perception, we think that at any moment we 'see' over a wider area and in more detail than we in fact do. Also, under very special circumstances we can show that our visual perception relies on detecting changes to a much greater extent than we are normally aware of. This is an important observation in avoiding an infinite regress of 'internal cameras looking at internal televisions'. It seems to us that we have a camera like view of the world, but in fact there are qualitative changes which transform the information. These changes do not just change high and low photon levels to high and low voltage levels, but re-organise the information through processes such as differencing and comparison. The observation becomes even more important in the context of our perceptions of how we think. It may seem to us that we have an awareness of our own thoughts as they occur and that this is somehow separate from the thoughts themselves. The separateness could, however, be illusory in the same sense that our apparent pixel perfect view of the visual world is an illusion. Dennett goes further and argues that even the consciousness of 'I' as a single person with time ordered past and present is something of an illusion. The mind is well capable of re-ordering (mental) events so that they appear to occur in an order that 'makes sense'. He reverses the well known simulation of parallel hardware on serial machines, by suggesting that we appear to have a serial existence that is actually implemented by mechanisms acting in parallel! Dennett puts forward the idea that our perception of events is in some ways analogous to a story that has multiple drafts. As the perceptions reverberate in the brain different units add detail, related information and judgements. Dennett also picks out a key mental pattern, that of asking ourselves internal questions and using these to guide our perceptions. This aspect to the argument provides a model for dreams and hallucinations, for in these instead of positing a 'script writer' and a 'script reader', one posits that the mechanisms for answering questions using external stimuli is off duty. Comment Dennett argues the case for not trusting one's perception of internal events so persuasively that one is tempted to entirely reject 'introspection' as a valid route into understanding the mind - though I and Dennett do not go this far. However, seen in the context of uncritical acceptance of introspection, expressing the counter arguments strongly is essential. I agree with Dennett on the difficulties of introspective enquiry. I find that when I try to understand processes in my own mind and express them in words, I keep returning to the most verbal thoughts. These are the kind of thought most accessible to my verbal understanding. The very process of expressing a thought in words changes it. An introspective inquiry needs to recognise this problem and attempt to gauge its importance. The notion of serial consciousness on a parallel framework is a very appealing one. For me the most crucial distinction this idea spotlights is between: How we actually arrive at some conclusion and... Our (reported) memory of how we did so. It seems that whilst our report is linear, the actual process was indeed a parallel one. I draw an analogy here with the adage: "History is always written by the winner". My feelings are that when I reconstruct the route by which I arrived at a particular conclusion, the ideas that did not materially help me reach that conclusion tend to get written out of the history. From a games theory perspective, it is as if my perception of my own thought is closer to the pruned tree of possibilities than it is to the tree of possibilities I actually explored. The question-answer pattern is certainly important. It's also a subdivision of function which has explanatory power for the questioner and the answerer could each be seen as having a much simpler mechanism than a complete mind. This is not really the case for the writer/reader subdivision. I agree with Dennett's idea that dreams reflect this pattern. Dennett's proposal that a question-answer mechanism operates in dreams is not incompatible with the idea that dreams have a valuable function . Dennett's description certainly frees us from a need for consciousness to be localised in one place in the brain. In his model it is inherently distributed.

Virtual Serial Machine

Consciousness Explained - Daniel C Dennett. ISBN 0-14-012867-0 (1991).

A crucial observation is that our perception of our own internal processes is not a reliable guide to how they work. For example, with our visual perception, we think that at any moment we 'see' over a wider area and in more detail than we in fact do. Also, under very special circumstances we can show that our visual perception relies on detecting changes to a much greater extent than we are normally aware of.

This is an important observation in avoiding an infinite regress of 'internal cameras looking at internal televisions'. It seems to us that we have a camera like view of the world, but in fact there are qualitative changes which transform the information. These changes do not just change high and low photon levels to high and low voltage levels, but re-organise the information through processes such as differencing and comparison.

The observation becomes even more important in the context of our perceptions of how we think. It may seem to us that we have an awareness of our own thoughts as they occur and that this is somehow separate from the thoughts themselves. The separateness could, however, be illusory in the same sense that our apparent pixel perfect view of the visual world is an illusion.

Dennett goes further and argues that even the consciousness of 'I' as a single person with time ordered past and present is something of an illusion. The mind is well capable of re-ordering (mental) events so that they appear to occur in an order that 'makes sense'. He reverses the well known simulation of parallel hardware on serial machines, by suggesting that we appear to have a serial existence that is actually implemented by mechanisms acting in parallel!

Dennett puts forward the idea that our perception of events is in some ways analogous to a story that has multiple drafts. As the perceptions reverberate in the brain different units add detail, related information and judgements.

Dennett also picks out a key mental pattern, that of asking ourselves internal questions and using these to guide our perceptions. This aspect to the argument provides a model for dreams and hallucinations, for in these instead of positing a 'script writer' and a 'script reader', one posits that the mechanisms for answering questions using external stimuli is off duty.

Comment
Dennett argues the case for not trusting one's perception of internal events so persuasively that one is tempted to entirely reject 'introspection' as a valid route into understanding the mind - though I and Dennett do not go this far. However, seen in the context of uncritical acceptance of introspection, expressing the counter arguments strongly is essential.
I agree with Dennett on the difficulties of introspective enquiry. I find that when I try to understand processes in my own mind and express them in words, I keep returning to the most verbal thoughts. These are the kind of thought most accessible to my verbal understanding. The very process of expressing a thought in words changes it. An introspective inquiry needs to recognise this problem and attempt to gauge its importance.
The notion of serial consciousness on a parallel framework is a very appealing one. For me the most crucial distinction this idea spotlights is between:

How we actually arrive at some conclusion and...
Our (reported) memory of how we did so.
It seems that whilst our report is linear, the actual process was indeed a parallel one. I draw an analogy here with the adage: "History is always written by the winner". My feelings are that when I reconstruct the route by which I arrived at a particular conclusion, the ideas that did not materially help me reach that conclusion tend to get written out of the history.
From a games theory perspective, it is as if my perception of my own thought is closer to the pruned tree of possibilities than it is to the tree of possibilities I actually explored.
The question-answer pattern is certainly important. It's also a subdivision of function which has explanatory power for the questioner and the answerer could each be seen as having a much simpler mechanism than a complete mind. This is not really the case for the writer/reader subdivision. I agree with Dennett's idea that dreams reflect this pattern.
Dennett's proposal that a question-answer mechanism operates in dreams is not incompatible with the idea that
dreams have a valuable function

.
Dennett's description certainly frees us from a need for consciousness to be localised in one place in the brain. In his model it is inherently distributed.

•Top•Down to Wolff•

Baldwin Effect Dennett mentions the 'Baldwin Effect' in describing how something as complex as the brain could evolve. It is an important theoretical idea. The essence is that it is 'easier' for evolution to discover a flexible system than a highly specialised inflexible one! The key is that an inflexible system has essentially one chance of being right. A flexible system can in one lifetime try a huge number of 'combinations' and lock on to the 'right' one when it finds it. Comment In my view 'The Baldwin Effect' gives a hint of the significance of multiple levels in feedback processes, particularly in evolution. It's one small aspect of a much larger whole. My article on evolution goes into more depth and shows how varied in character the levels of feedback in evolution can be. As well as evolution, the 'Baldwin Effect' also applies to para-evolutionary situations. For example in the immune response to a new antigen there is an initial response with large general purpose molecules, which are specialised to the particular antigen. As they hone their technique, the body switches over to making much smaller, cheaper to manufacture and more specific molecules.

Baldwin Effect

Dennett mentions the 'Baldwin Effect' in describing how something as complex as the brain could evolve. It is an important theoretical idea. The essence is that it is 'easier' for evolution to discover a flexible system than a highly specialised inflexible one!

The key is that an inflexible system has essentially one chance of being right. A flexible system can in one lifetime try a huge number of 'combinations' and lock on to the 'right' one when it finds it.

Comment
In my view 'The Baldwin Effect' gives a hint of the significance of multiple levels in feedback processes, particularly in evolution. It's one small aspect of a much larger whole. My article on evolution goes into more depth and shows how varied in character the levels of feedback in evolution can be.
As well as evolution, the 'Baldwin Effect' also applies to para-evolutionary situations. For example in the immune response to a new antigen there is an initial response with large general purpose molecules, which are specialised to the particular antigen. As they hone their technique, the body switches over to making much smaller, cheaper to manufacture and more specific molecules.

•Top•Down to Pinker•

Cognition as Compression

The idea of Cognition as Compression (and related ideas of Computing as Compression ) are described by researcher Gerry Wolff on the .sees. website. Two papers on that website (1) describing a method , (2) giving examples , give Dr Wolff's presentation of a model for parsing language:

Information Compression by Multiple Alignment, Unification and Search (ICMAUS).

With ICMAUS compression, sentences after compression display the language structure. The compressions are parsings of the sentences.

Parsing is an application of the model, for the underlying model is more general. It is not restricted to grammatical information.

Some neat twists with the ICMAUS model include:

The same model can be used for language production. Both parsing and production can be seen as compression processes - and both use the same engine.
The model gracefully accommodates imperfect and ambiguous information. These appear as lower quality alignments and alternative alignments.
Multiple alignments provide a representation which does not force a tree-like structuring of information (whilst still allowing it). This is an area where traditional recursive descent parsing has difficulty since without special purpose fix-ups recursion 'forces' a tree-like analysis.

A planned stage in the research is development of radically new computer architectures to support the framework.

Comment
These ideas are exciting and produce results. The argument that understanding information and compressing it are closely related is compelling. Both require abstracting pattern from raw information. Both require relating new information to prior information

Key ICMAUS Compression Properties
Seen purely as a compression technique two key aspects of the ICMAUS approach are:

Ability to compress non-contiguous runs of information.
Compression technique that is not confined to the raw information. It works simultaneously on the raw information and 'compressed fragments'.
To me the essential aspect of ICMAUS are these two properties rather than the multiple alignment representation which achieves them.

Recursion?
Under the hood, I think there is a closer connection between recursive descent parsing and ICMAUS parsing than at first appears. An ICMAUS multiple alignment can represent a traditional parse precisely. As I see it, in ICMAUS recursion is buried in there in the foundations. ICMAUS relies on multiple alignment which in turn (in my own research) relies on dynamic programming which in turn is an upgrade of recursion. Elsewhere I've shown that any recursive algorithm can be upgraded to dynamic programming. It looks to me as if ICMAUS, with its non treelike structures, takes 'recursion' further still. You can read this algorithmic connection either as:

ICMAUS isn't such a novel method as it seems OR read it as I do...
An ICMAUS framework has the potential to transform a wide range of algorithms - including almost any search algorithm based on recursion.

Hi and Low levels
From another angle, the ICMAUS approach brings to language processing the same high level benefits that the relational approach brought to databases. It gives a very clean abstract framework. This isn't an accident for the joins between sequence fragments are in some ways like joins of database tables. The formal presentation in both cases can show regular structures and can keep a morass of pointers - the study of which is unilluminating - hidden from view.
As the research moves on to its next stage, production of new architectures, I expect that prototyping that addresses the lower level issues such as speed will increasingly dominate.
As I see it the low level speed aspect is more than a nuts-and-bolts enterprise for it is here that the extent to which the compression algorithm is hardwired and the extent to which it is itself open to change is decided. That decision in turn determines what the system will be able to learn.

•Top•Down to Postscript•

The Language Instinct The Language Instinct - Steve Pinker, ISBN 0-14-017529-6, (1994) Pinker's central thesis is that language acquisition is so amazing that we must posit a very special mechanism for it. The argument gives specific details of linguistic complexity, shows the astonishing speed of acquisition, points to commonality of structures (and errors) in different languages, and cites precedents for unique attributes in the animal kingdom. A subsidiary discussion is of the difference between grammar rules (which are infinitely nestable) and our actual parsing process, which can run out of steam very rapidly (nestings three deep) when the same rule is used recursively. In 'The Language Instinct' there are many interesting examples of 'quirks' of language which show that grammatical parse trees (on their own) are too pure an approach. Actual language use is far more diverse and irregular than the parse tree analysis suggests. Comment I think Pinker's claims that there must be special mechanisms for language acquisition - on the basis of the speed with which we learn new words - is overstated. It's the emphasis on 'special' that I am uncomfortable with. I think we learn other things rapidly too, so could it not be that we have general mechanisms for acquiring knowledge rapidly and that language learning is just a minor variant of this? I liked the discussion contrasting formal grammar and a parser for it. It allows for a clean model (the grammar) with any limitations being implementation details (the parser). As a programmer I can interpret the observed human parsing limit on nesting depth as 'storage-with-rule' being used not 'storage-on-stack', for this 'explains' the difficulty of dangling resolutions where the same rule is re-used. The overall thrust of the grammar / parser discussion points convincingly to something less pure than a grammar being the 'real' underlying structure. From the many diverse examples, an overall message I took home is that any programmer working on language who chooses not to look at examples of subtlety in language will end up with a parser limited to artificial examples. The subtleties actually look central, not things that can be bolted on later as a quick fix.

The Language Instinct

The Language Instinct - Steve Pinker, ISBN 0-14-017529-6, (1994)

Pinker's central thesis is that language acquisition is so amazing that we must posit a very special mechanism for it. The argument gives specific details of linguistic complexity, shows the astonishing speed of acquisition, points to commonality of structures (and errors) in different languages, and cites precedents for unique attributes in the animal kingdom.

A subsidiary discussion is of the difference between grammar rules (which are infinitely nestable) and our actual parsing process, which can run out of steam very rapidly (nestings three deep) when the same rule is used recursively. In 'The Language Instinct' there are many interesting examples of 'quirks' of language which show that grammatical parse trees (on their own) are too pure an approach. Actual language use is far more diverse and irregular than the parse tree analysis suggests.

Comment
I think Pinker's claims that there must be special mechanisms for language acquisition - on the basis of the speed with which we learn new words - is overstated. It's the emphasis on 'special' that I am uncomfortable with. I think we learn other things rapidly too, so could it not be that we have general mechanisms for acquiring knowledge rapidly and that language learning is just a minor variant of this?
I liked the discussion contrasting formal grammar and a parser for it. It allows for a clean model (the grammar) with any limitations being implementation details (the parser). As a programmer I can interpret the observed human parsing limit on nesting depth as 'storage-with-rule' being used not 'storage-on-stack', for this 'explains' the difficulty of dangling resolutions where the same rule is re-used. The overall thrust of the grammar / parser discussion points convincingly to something less pure than a grammar being the 'real' underlying structure.
From the many diverse examples, an overall message I took home is that any programmer working on language who chooses not to look at examples of subtlety in language will end up with a parser limited to artificial examples. The subtleties actually look central, not things that can be bolted on later as a quick fix.

•Top•Down to Links•

PostScript Various authors have come at the problem of 'the mind' from different angles, from neurobiology, from an interest in what awareness is, from an AI stance. The different ideas are not mutually inconsistent. Often one 'explanation' does not over rule another - though it is easy to think that because one explanation 'is right' another must be wrong. For me some of the more interesting connections include: Dennett focuses on the necessity of recoding information. Wolff picks out the need for more economical recodings. From both viewpoints, the mechanism that transforms the information must abstract some essential quality, otherwise our model of the brain is not doing anything useful. An interplay between serial results (e.g. language) and parallel mechanisms. Dennett takes this from the point of view of our perception of consciousness and sensation information coming to us in parallel, Wolff from the point of view of algorithmic complexity - you need parallelism for speed. Dennett argues that consciousness doesn't happen at one location; By contrast, Luria's research was aimed at localising functions; both reject a crude and unsophisticated phrenological location-function correspondence; both argue that spatially distributed systems are what make up the brain's functional units. Dennett picks out underlying mechanisms, e.g. he sees a general mechanism of 'internal dialogue' (questions and answers) as being related to a general mechanism in perception of directed attention; Pinker meanwhile explores the diversity of properties of language and suggests they reflect distinct mechanisms. It as if examining a tree, Dennett looks at the branch-points and Pinker at the leaves. Interplay between very clean abstract systems (e.g. grammar) and less clean details (e.g. actual language). Pinker debunks pure recursive grammars; Wolff devises a scheme that handles non-recursive structure and imperfect data. Each person brings a different focus. Someday these different viewpoints may come even closer together, old ideas will be discarded and the problems will move to new domains. As I see it there is already some consensus, albeit hidden, in the different lines researchers in different fields are taking.

PostScript

Various authors have come at the problem of 'the mind' from different angles, from neurobiology, from an interest in what awareness is, from an AI stance.

The different ideas are not mutually inconsistent. Often one 'explanation' does not over rule another - though it is easy to think that because one explanation 'is right' another must be wrong.

For me some of the more interesting connections include:

Dennett focuses on the necessity of recoding information. Wolff picks out the need for more economical recodings. From both viewpoints, the mechanism that transforms the information must abstract some essential quality, otherwise our model of the brain is not doing anything useful.
An interplay between serial results (e.g. language) and parallel mechanisms. Dennett takes this from the point of view of our perception of consciousness and sensation information coming to us in parallel, Wolff from the point of view of algorithmic complexity - you need parallelism for speed.
Dennett argues that consciousness doesn't happen at one location; By contrast, Luria's research was aimed at localising functions; both reject a crude and unsophisticated phrenological location-function correspondence; both argue that spatially distributed systems are what make up the brain's functional units.
Dennett picks out underlying mechanisms, e.g. he sees a general mechanism of 'internal dialogue' (questions and answers) as being related to a general mechanism in perception of directed attention; Pinker meanwhile explores the diversity of properties of language and suggests they reflect distinct mechanisms. It as if examining a tree, Dennett looks at the branch-points and Pinker at the leaves.
Interplay between very clean abstract systems (e.g. grammar) and less clean details (e.g. actual language). Pinker debunks pure recursive grammars; Wolff devises a scheme that handles non-recursive structure and imperfect data.

Each person brings a different focus. Someday these different viewpoints may come even closer together, old ideas will be discarded and the problems will move to new domains. As I see it there is already some consensus, albeit hidden, in the different lines researchers in different fields are taking.

•Top•

Mind Links (From Min-Review):

Cognition as Compression: Click here to jump back to that section on this page where there are links to Gerry Wolff's site.

•Catalase Website Index •James •Techno •Thought •Thought Mini-Review •

"Approaches to Mind - A Selective Review" page last updated 5-July-2003