Here’s a topic of which I believe more people, especially AI system desigers, should be aware: David Marr’s three-level design hierarchy. I believe mere understanding of this hierarchy can answer a lot of questions people often ask, such as “What is the future of neural networks?” and “Where should we focus for a breakthrough in AI?”

Somehow when I bought Marr’s book “Vision” years ago, I missed the main concept he was trying to impart on this subject. I think my oversight occurred for several reasons: Marr’s verbose description drags on for several pages, there was no clear example that spanned all three layers, the obvious 3-level illustration that I would expect to see was not there, he never used the word “hierarchy” anywhere so I never understood the relationships between those three layers, and the names he gave to those layers were misleading. It wasn’t until I read Dennett’s book “Brainchildren” that discussed Marr’s layers that I finally understood what Marr had intended, and its importance. After all, if multiple textbooks nowadays mention Marr’s hierarchy, it must be important.

Here are the three proposed levels as Marr described them, though I inserted the numbers and shortened the names of those layers to match Dennett’s description:

(1) Computational

What is the goal of the computation, why is it appropriate, and what is the logic of the strategy by which it can be carried out?

(2) Algorithmic

How can this computational theory be implemented? In particular, what is the representation for the input and output, and what is the algorithm for the transformation?

(3) Physical

How can the representation and algorithm be realized physically?

(“Vision: A Computational Investigation into the Human Representation and Processing of Visual Information”, David Marr, 1982, page 25)

Dennett discusses the misleading names that Marr used for these levels, and finds some flaws with Marr’s hierarchy…

The problem is that we do not yet have many clear ideas about what
the functions of such systems must be—what they must be able to do.
This setting of the problem has been forcefully developed by David
Marr (1982) in his methodological reflections on his work on vision.
He distinguishes three levels of analysis. The highest level, which he
rather misleadingly calls computational, is in fact not at all concerned
with computational processes, but strictly (and more abstractly) with
the question of what function the system is serving—or,
more formally, with what function in the mathematical sense it must
(somehow or other) “compute.” Recalling Chomsky’s earlier version
of the same division of labor, we can say that Marr’s computational
level is supposed to yield a formal and rigorous specification of a sys-
tem’s competence—“given an element in the set of x’s it yields an
element in the set of y’s according to the following rules”—while re-
maining silent or neutral about implementation of performance. Marr’s
second level down is the algorithmic level, which does specify the com-
putational processes but remains neutral (as neutral as possible) about
the hardware, which is described at the bottom level.

Marr’s claim is that until we get a clear and precise understanding
of the activity of a system at the highest, “computational” level, we
cannot properly address detailed questions at the lower levels, or inter-
pret such data as we may already have about processes implementing
those lower levels. Moreover, he insists, if you have a seriously mis-
taken view about what the computational-level description of your sys-
tem is (as all earlier theorists of vision did, in his view), your attempts
to theorize at lower levels will be confounded by spurious artifactual
problems.
(“Brainchildren”, Daniel C. Dennett, 1998, pages 230, 231)

Dinsmore also discusses these names of these levels, and implies that other authors, including earlier authors, have proposed alternative names for these same layers, which underscores the significance of this hierarchy:

Levels of Description

The symbolic paradigm traditionally sees cognition in a broad sense as
a sandwich. The top slice of bread is variously called the knowledge
level (Newell 1982), the intentional (Dennett 1978) or the computa-
tional level (Marr 1982). The bottom slice of bread is the level of
hardware implementation (Marr 1982) or the physical level (Dennett
1978). The salami itself is the symbol (Newell 1982), design (Dennett
1978) or representations and algorithms (Marr 1982) level.
(“Brainchildren”, Daniel C. Dennett, 1998, page 12)

I often claim that the prime area for productive research in AI is in knowledge representation (KR). My belief is partly based on this 3-level hierarchy. Clearly the bottom layer is of limited insight since the physical implementation of any algorithm could be done in many ways, say on a digital computer versus an analog computer, or on a classical computer versus a quantum computer. If you pried open a television set, for example, the appearance of the raw circuitry would not help you much to understand generally how the device works. For that understanding, you need one of the top two layers of the hierarchy.

Marr explicitly states that representation occurs at the middle (Algorithmic) level. Why then do I claim that the important layer is the middle layer instead of the top (Computational) level? Multiple reasons: (1) We often already know what we want our system to do, such as recognize a face or find the shortest path, so specifying that at the top level is unlikely to give us new insights into how to program that. (2) The description of our goal at the top level (e.g., play tic-tac-toe, or rotate an image) often implicitly requires describing the objects on which are being operated (e.g., a 3x3 square grid, or a color photograph) in order for the goal to even make sense, so to get to the top level we often have to know what is in the middle level anyway. (3) The middle level is where data structures lie, which is what we are already used to thinking about with respect to computer languages and programming paradigms, such as whether to describe the problem with lists (as in LISP) or objects (as in C++) or numbers (as in C) or logic (as in PROLOG).

Now it’s easy to see why artificial neural networks (ANNs) have never really been successful, and why they probably won’t ever be successful per se: ANNs are at the bottom/implementation level of abstaction, which is the least relevant level. It doesn’t matter much, other than for efficiency issues, whether that level is implemented with neurons, digital circuits, analog circuits, an abacus, protein molecules, Tinkertoys, or something else: what the user will see and care about in a system as a whole is the high-level phenomena at the macro level. As far as the typical user and the public is concerned, the nuts and bolts at the bottom level are for engineers who specialize in such details that would give most other people headaches just to think about.

Accordingly, it’s also easier to see where to focus on a breakthrough in AI: look in one of the top two levels.

This brings us to the third level, that of the device in which the process
is to be realized physically. The important point here is that, once again,
the same algorithm may be implemented in quite different technologies.
The child who methodically adds two numbers from right to left, carrying
a digit when necessary, may be using the same algorithm that is imple-
mented by the wires and transistors of the cash register in the neighbor-
hood supermarket, but the physical realization of the algorithm is quite
different in these two cases. Another example: Many people have written
computer programs to play tic-tac-toe, and there is a more or less standard
algorithm that cannot lose. This algorithm has in fact been implemented
by W. D. Hilis and B. Silverman in a quite different technology, in a com-
puter made out of Tinkertoys, a children’s wooden building set. The whole
monstrously ungainly engine, which actually works, currently resides in a
museum at the University of Missouri in St. Louis.
(“Vision: A Computational Investigation into the Human Representation and Processing of Visual Information”, David Marr, 1982, page 24)

With Marr’s hierarchy, it’s also easier to see some of the influences on my own definition of “intelligence” (“Intelligence with respect to a given goal is the ability to perform efficient, adaptive processing of real-world data for attaining that goal.”). Note how Marr uses almost all the same concepts my definition does, and even many of the words verbatim:

(1) goal

What is the goal of the computation, why is it appropriate, and what is the logic of the strategy by which it can be carried out?
(“Vision: A Computational Investigation into the Human Representation and Processing of Visual Information”, David Marr, 1982, page 25)

(2) efficient

And
third, even for a given fixed representation, there are often several possible
algorithms for carrying out the same process. Which one is chosen will
usually depend on any particularly desirable or undesirable characteristics
that the algorithms may have; for example, one algorithm may be such
more efficient than another, or another may be slightly less efficient but
more robust (that is, less sensitive to slight inaccuracies in the data on
which it must run).
(“Vision: A Computational Investigation into the Human Representation and Processing of Visual Information”, David Marr, 1982, pages 23, 24)

(3) processing

It is clear that human vision is much more complex than this, although
it may well incorporate subsystems not unlike the fly’s to help with specific
and rather low-level tasks like the control of pursuit eye movements. Never-
theless, as Poggio and Reichardt have shown, even these simple systems
can be understood in the same sort of way, as information processing tasks.
(“Vision: A Computational Investigation into the Human Representation and Processing of Visual Information”, David Marr, 1982, page 35)

(4) real-world data

The hope that lay behind this work was, of course, that once the toy
world of white blocks had been understood, the solutions found there
could be generalized, providing the basis for attacking the more complex
problems posted by a richer visual environment. Unfortunately, this turned
out not to be so. For the roots of the approach that was eventually suc-
cessful, we have to look at the third kind of development that was taking
place then.
(“Vision: A Computational Investigation into the Human Representation and Processing of Visual Information”, David Marr, 1982, page 17)

Accidental corroboration such as Marr’s is nice because it suggests my definition is on track, which in turn suggests that my research directions are also on track.

Currently I am trying to train myself to always think in terms of Marr’s hieararchy when designing AI systems. I spent so many years dealing with ANNs that I still tend to think in terms of how to implement designs with ANNs when in fact there may be much better hardware implementations that involve neither digital circuits or ANNs.

Since Marr omitted some good, clear-cut examples identified by each level,
I’ll supply such examples, some of his, some of mine, laid out more clearly by level.
(I’m out of space for this thread: I’ll post the list if someone responds.)

These three layers sound like the first three steps of classic software engineering: specification, design, implementation. The final step before the entire process repeats is testing. The whole process is guided by goals set at the start of the project.

Some succinct examples might illustrate how Marr’s hierarchy is different from the usual software life cycle.

Andrew Smith - Jul 26, 2013:

These three layers sound like the first three steps of classic software engineering: specification, design, implementation.

I think the main difference is that when a person is designing a program, they already know which software paradigm they’re going to use, and it’s usually assumed they’re going to use a digital computer, usually a single-processor digital computer, therefore the entire software development cycle exists at a single Marr level, namely at the middle/Algorithmic level. Marr’s 3-level hierarchy is aimed at the general design process of an entire vision system, from the overall goal to the nuts-and-bolts hardware level, therefore his hierarchy can apply to some very exotic architectures—the kind that really interest me. I think it’s important to realize that this process applies to any AI systems, though, especially since there are some people who believe that vision is a key aspect of strong artificial intelligence.

I’m still getting used to Marr’s “loosely coupled” hierarchy, so if you detect any errors in my examples, or if you can think of other good examples, or if you detect any flaws in my understanding of this hierarchy, please let us know.

Here are some examples that I believe are consistent with Marr’s intent, though Marr was more focused on mathematical operations at the top level than I am. (I’m planning another thread or two on that topic alone!)

()
Computational level: add two integers
Algorithmic level: align the integers, use carries from right to left
Physical level: pencil and paper
()
Computational level: add two integers
Algorithmic level: align the integers, use carries from right to left
Physical level: cash register
()
Computational level: play tic-tac-toe
Algorithmic level: 3x3 square grid marked with Xs and Os
Physical level: digital computer
()
Computational level: play tic-tac-toe
Algorithmic level: 3x3 square grid marked with Xs and Os
Physical level: Tinkertoys
()
Computational level: rotate a photograph
Algorithmic level: image, using trigonometric description of rotation
Physical level: digital computer
()
Computational level: rotate a photograph
Algorithmic level: image, using physical motion of the image substrate
Physical level: human hands
()
Computational level: sort a list of positive numbers
Algorithmic level: bubble sort, using magnitudes verbatim
Physical level: digital computer
()
Computational level: sort a list of positive numbers
Algorithmic level: balls with diameters corresponding to the magnitudes of the numbers
Physical level: rolling tracks that cause the smaller balls to fall through first

  Since the hierarchy is only a rough guideline, there exist flaws if you consider it too exactly. Dennett notes that some mechanisms can span all the levels, and I believe there are probably sublevels or subdivisions within levels, such as which programming language level (e.g., high level versus assembly language) is being used.
  I hope Marr’s design layout will be of use to people here who are designing intelligent systems of any kind.

For example, if one chooses the Arabic numeral
representation, it is easy to discover whether a number is a power of 10
but difficult to discover whether it is a power of 2. If one chooses the binary
representation, the situation is reversed. Thus, there is a trade-off; any
particular representation makes certain information explicit at the expense
of information that is pushed into the background and may be quite hard
to recover.
  This issue is important because how information is represented can
greatly affect how easy it is to do different things with it. This is evident
even from our numbers example: It is easy to add, to subtract, and even to
multiply if the Arabic or binary representations are used, but it is not at all
easy to do these things—especially multiplication—with Roman numerals.
This is a key reason why the Roman culture failed to develop mathematics
in the way the earlier Arabic cultures had.
  An analogous problem faces computer engineers today. Electronic
technology is much more suited to a binary number system than to the
conventional base 10 system, yet humans supply their data and require the
results in base 10. The design decision facing the engineer, therefore, is,
Should one pay the cost of conversion into base 2, carry out the arithmetic
in a binary representation, and then convert back into decimal numbers
on output; or should one sacrifice efficiency of circuitry to carry out oper-
ations directly in a decimal representation?
(“Vision: A Computational Investigation into the Human Representation and Processing of Visual Information”, David Marr, 1982, page 21)

—————

ERRATA

(Sorry: Next time I’ll use a spell checker!)

“desigers” should be “designers”

“abstaction” should be “abstraction”.

The supposed Dinsmore quote is really by Dinsmore, not by Dennett:

Levels of Description

The symbolic paradigm traditionally sees cognition in a broad sense as
a sandwich. The top slice of bread is variously called the knowledge
level (Newell 1982), the intentional (Dennett 1978) or the computa-
tional level (Marr 1982). The bottom slice of bread is the level of
hardware implementation (Marr 1982) or the physical level (Dennett
1978). The salami itself is the symbol (Newell 1982), design (Dennett
1978) or representations and algorithms (Marr 1982) level.
(“Partitioned Representations: A Study in Mental Representation, Language Understanding and Linguistic Structure”, John Dinsmore, 1991, page 12)

Another reason I believe the key level for discovery is the middle layer:

(4) Just knowing the objects that need to be manipulated to solve the problem can make a solution much clearer. An analogy might be that if a monkey were given a long weighted string, that tool/object would likely suggest one solution to the Monkey and Banana Problem (http://en.wikipedia.org/wiki/Monkey_and_banana_problem) that would not have been considered otherwise.

I believe in multi-level design, but I think Marr’s labels (Computational, Algorithmic, Physical) lead to confusion, especially in AI development. Each of these are trade-offs in the performance of a system. Good Algorithms trump computational and physical constraints.

Sometimes we confuse:
Can it be done? vs What is the optimal way to do it?

With AGI, we are still working of the first section. Once it has been done, optimizing is easier because we have a working specification/prototype. As Andrew points out; specification, design, implementation, is a standard software/hardware development paradigm. Where this type of approach leads to trouble is when it is difficult/impossible to create a specification, or when the design is locked into a specific architecture (hardware/software).

In “systems” design, I often look at the concept/goals of the system starting from a very simplified structure:
Input -> Process -> Output

In many cases Inputs and desired Outputs are well defined. Sometimes complex Inputs/Outputs can be transformed to simpler more stable ones.
Input:
Sound->Speech Recognition->Text
Output:
Text->TTS->Sound

The “Process” is the black box we are architecting. In many ways, the best analogy is that we are “architects”. The design depends on what neighborhood and aesthetics we are trying to satisfy. The actual construction could be done by a variety of craftsmen if we were able to give them the correct blueprints.

“Process” analysis and design can be made more digestible by breaking it down further into subsystems of smaller (Input -> Process -> Output) chunks.

Other “system” factors to consider:
flexibility
configurability
modularity

Speed, or optimization will handle itself via Moore’s law. I would suggest with the cloud and supercomputers we already have enough compute power for AGI. What we don’t have is a good blueprint.

Computational level: add two integers
Algorithmic level: align the integers, use carries from right to left
Physical level: pencil and paper

If I look at this toy problem with an eye toward the (Input -> Process -> Output) chunk, it highlights other system design issues.

INPUT:
Operation: add
Symbol 1: integer
Symbol 2: integer
(here we hit our first problem, AI grounding. What is an integer, “00000010, two, 2”? What does “add” mean? Is “Add two plus two.” the same as “2+2?”? How do we filter this input from all other inputs to trigger this operation?)

Process (black box):
Intelligent entity with instructions to transform input to output
Could be human with pencil and paper, abacus, calculator, software/hardware/cloud
The implementation of this implies trade-offs.
“2+2?” could be directly mapped as a string to “4”. (Maybe that is what humans do as we learn addition and multiplication tables when we are young.)
Alternatively it could be:
“2+2?”->Math.Sum(2,2)
Both approaches would solve the problem but they represent 2 different philosophies in system design.

Output (goal):
Sum of Symbol 1 and Symbol 2 (add two integers)

With Marr’s hierarchy, it’s also easier to see some of the influences on my own definition of “intelligence” (“Intelligence with respect to a given goal is the ability to perform efficient, adaptive processing of real-world data for attaining that goal.”).

I would submit that efficiency is not required for intelligence. A brute force approach to many tasks may not be efficient, but still may get the job done. There are a variety of ways to look up a key:value pair, each will return a correct value. Doing it the most efficient way may not be necessary. I am also not sure about your definition of “adaptive processing”. Maybe you could go into your thoughts a bit more.

Merlin - Jul 28, 2013:

Good Algorithms trump computational and physical constraints.

Well, I’m sure one of us here could think of an example where extreme physical constraints guaranteed that no algorithm could ever run on that hardware in any reasonable time, but I get your point. I pretty much agree with you on your other points, too, and I’ve been thinking about the symbol grounding problem quite a bit recently myself.

Merlin - Jul 28, 2013:

I would submit that efficiency is not required for intelligence. A brute force approach to many tasks may not be efficient, but still may get the job done.

To me, that just means you set the efficiency parameter in my definition to almost zero, which is fine: my definition allows for that and just claims that the resulting intelligence diminishes but does not disappear until efficiency hits zero. Typically the efficiency to which I’m referring in my definition is time efficiency, basically speed per resources expended, but it could also apply to space efficiency, energy efficiency, complexity, or some other type of efficiency.

Expressiveness and efficiency of reasoning are listed in virtually
any textbook on artificial intelligence as the primary factors that
motivate specific models of knowledge representation. Usually left
implicit in such discussion is rationality.
(“Partitioned Representations: A Study in Mental Representation, Language Understanding and Linguistic Structure”, John Dinsmore, 1991, page 37)

Further, the societal concept of intelligence contains the notion that
speed of performance is at least sometimes important. It is recognized (at
least in most advanced societies) that time for solving problems is not
infinite, and the person who can solve a problem more quickly than another
is regarded more highly, and thus possibly as more capable and “intelligent.”
(“What Is Intelligence?”, John B. Carroll, in “What Is Intelligence?: Contemporary Viewpoints on Its Nature and Definition”, Robert J. Sternberg & Douglas K. Detterman, eds., 1986, page 52)

It is difficult not to
conclude that these data support some such notion of intelligence as
being based on speed of cortical processing of information.
(“Intelligence: A New Look”, Hans J. Eysenck, 1998, page 58)

All ECTs by definition are of so simple a kind
that even mentally defective children can solve the “problem” fault-
lessly; the only thing that differentiates bright and dull is the speed
with which the task is carried out. That was Galton’s prediction, and
the facts certainly bear out this prediction remarkably well. Further-
more, speed of mental processing so measured correlates equally well
with all types of IQ tests, and best with those having highest load-
ing on g (i.e., those which are the best measures of g) this would be
difficult to explain on any other grounds.
  Speed of mental functioning is clearly very relevant to IQ testing,
although I shall argue in the next chapter that there is an even more
fundamental biological variable that underlies such speed measures.
But however that may be, ECTs have an important bearing on thie Bi-
net-Galton controversy. Apparently abstract ability, reasoning, learn-
ing and memory are all dependent on speed of cortical functioning;
that is an important lesson to learn.
(“Intelligence: A New Look”, Hans J. Eysenck, 1998, page 59)

The need for focusing may be argued on the grounds of a limited capacity that is
ordinarily available for information processing, as Mesulam (1985) points out in the
context of human attention: “If the brain had infinite capacity for information processing,
there would be little need for attentional mechanisms.” From this quote, we may infer that
the use of focusing provides a mechanism for a more efficient utilization of information-
processing resources.
(“Neural Networks: A Comprehensive Foundation”, Simon Haykin, 1994, page 612)

    TABLE 1.1

Frequencies of Attributes that Contributors Used to Define Intelligence in 1986 and 1921.

                                            1986     1921
                                            No.%      No. %

1. Adaptation, in order to meet the demands of the
  environment effectively                           3   13     4   29
2. Elementary processes (perception, sensation, attention)      5   21     3   21
3. Metacognition (knowledge about cognition)              4   17     1   7
4. Executive processes                             6   25     1   7
5. Interaction of processes and knowledge                 4   17     0   0
6. Higher level components (abstract reasoning, represent-
  ation, problem solving, decision making)                12 50     8   57
7. Knowledge                                     5   21     1   7
8. Ability to learn                                4   17     4   29
9. Physiological mechanisms                           2   8     4   29
10. Discrete set of abilities (e.g., spatial, verbal, auditory)  4   17     1   7
11. Speed of mental processing                        3   13     2   14
12. Automated performance                           3   13     0   0
13. g                                         4   17     2   14
14. Real-world manifestations (social, practical, tacit)      2   8     0   0
15. That which is valued by culture                     7   29     0   0
16. Not easily definable, not one construct               4   17     2   14
17. A field of scholarship                           1   4     0   0
18. Capacities prewired at birth                       3   13     1   7
19. Emotional, motivational constructs                   1   4     1   7
20. Restricted to academic/cognitive abilities             2   8     2   14
21. Individual differences in mental competence             1   4     0   0
22. Generation of environment based on genetic programming     1   4     0   0
23. Ability to deal with novelty                       1   4     1   7
24. Mental playfulness                             1   4     0   0
25. Only important in its predictive value                 0   0     1   7
26. Inhibitive capacity                             0   0     1   7
27. Overt behavioral manifestation (effective/successful
  responses)                                  5   21     3   21

Source: Sternberg, 1990, p. 50.
(“The General Unified Theory of Intelligence: Its Central Conceptions and Specific Application to Domains of Cognitive Science”, Morton Wagman, 1997, page 13)

Merlin - Jul 28, 2013:

I am also not sure about your definition of “adaptive processing”. Maybe you could go into your thoughts a bit more.

I am using the word “adaptive” to mean “able to learn” or “having a memory”. Maybe my choice of that word is misleading: if so, please let me know. See the above Table 1.1 (sorry about the table text being out of alignment) for confirmation that both adaptation/learning and speed are commonly mentioned attributes in definitions of intelligence.

Here is an example of a (biological) system that is *not* adaptive, and the result is stupidity of such a degree that it would actually be humorous to watch a video of this experiment across all 40 trials…

By tweaking and keeping what works, evolution builds up complex behaviors. Some
of these are clearly instinctive (inherited and showing no sign of conscious under-
standing). Consider, for example, the egg-laying behavior of the wasp Sphex. This
wasp goes through an elaborate series of “subgoals” in laying its eggs. It builds a
burrow, finds and paralyzes a cricket, drags it to the burrow, checks the burrow,
places the cricket in the burrow, lays eggs in the burrow, and finally seals the burrow.
The wasp does not seem to “understand” the big picture of what it is doing. If the
cricket is moved a few inches away while the wasp is checking the burrow, the wasp
will drag the cricket back to the threshold and check again. This process was experi-
mentally observed on one occasion to have been repeated 40 times. Seemingly the
wasp is engaged in programmed behavior with no memory that it had just checked
the burrow. But at a minimum, the program is long and involved. Some wasps also
use tools: stones to tamp down soil to seal the burrow. But all these behaviors seem
innate and inflexible: change something, and the wasps don’t know how to react, or
else they simply go back to the previous step in their program and continue running
it (Franklin 1995).
(“What Is Thought?”, Eric B. Baum, 2004, page 354)

Anyway, my larger point was that Marr’s 3-level hierarchy seems to aid understanding of intelligent systems in many ways, whether dealing with vision, definitions, future of specific AI systems, valuable areas in which to focus, or other.

I agree that when measuring two systems that produce comparable results, the one that works faster or uses fewer resources is thought of as better (more intelligent). But if one system provides a correct response and the other provides an incorrect one, the first one will be thought of as more intelligent regardless of the timing. My fear is that by placing efficiency constraints on a system where the optimal algorithms are unknown, we may be missing important paths to the ultimate goal. Acceptable solutions may ultimately take the form of a S curve, but in the interim it would still be beneficial if we were able to produce solutions with suboptimal efficiency.

Speed does matter and does create the perception of intelligence. In the past few years, my bot was able to respond virtually instantaneously, compared to other bots that would take seconds or minutes to respond to an input. This has drawn many positive comments on how smart it is. On the other hand, although it can respond faster than any human, no one would say it is more intelligent (except when it comes to math).

I am using the word “adaptive” to mean “able to learn” or “having a memory”. Maybe my choice of that word is misleading: if so, please let me know.

I tend to view this in a more granular fashion, separating the ability to store and retrieve “memories” from planning, creativity, and extrapolating to unseen situations in the environment.

One theory of the mind is that is the connections that are important. The connectome may be as important as how individual neurons operate when trying to model the brain.

Sebastian Seung is mapping a massively ambitious new model of the brain that focuses on the connections between each neuron. He calls it our “connectome,” and it’s as individual as our genome—and understanding it could open a new way to understand our brains and our minds.

Sebastian Seung is a leader in the new field of connectomics, currently the hottest space in neuroscience, which studies, in once-impossible detail, the wiring of the brain.

Ted Talk on connectomics

Moore’s law will allow us to emulate the functions of the brain and allow us to test our theories.

Henry Markram is director of Blue Brain, a supercomputing project that can model components of the mammalian brain to precise cellular detail—and simulate their activity in 3D. Soon he’ll simulate a whole rat brain in real time

Ted talk on emulating a brain

Merlin - Jul 29, 2013:

I tend to view this in a more granular fashion, separating [1] the ability to store and retrieve “memories” from [2] planning, creativity, and extrapolating to unseen situations in the environment.

Hmm. I was worried about that. OK, let’s try this definition: “Intelligence with respect to a given goal is the ability to learn and efficiently process real-world data for attaining that goal.” You and CR keep me on my toes. That’s good.

Merlin - Jul 29, 2013:

Moore’s law will allow us to emulate the functions of the brain and allow us to test our theories.

My main complaint about waiting for Moore’s Law to catch up to the speed I want is that biological evolution probably didn’t work according to Moore’s Law, since all animals have been using the same types of neurons, animals needed to compete for millennia, and during that time evolution hit upon some method of data processing that still beats man’s 10^3 times faster circuits on real-world tasks. As Marvin Minsky said in an interview, we already have enough computer speed; what we need to know is only how to process that data in the right way.

Thanks for the links to Ted talks. I haven’t yet had a chance to watch them.

Some afterthoughts: (1) I don’t think Marr explicitly said that his hierarchy was for design, though that was implied at times. His hierarchy might have been mostly intended for analysis and understanding instead, since he reviewed the history of analyzing biological visual systems as he motivated that hierarchy. It works equally well either way, of course. (2) Another way that Marr’s hierarchical layers could be interpreted is, from top to bottom: idea, software, hardware. (3) Another book I read, probably Hans Moravec’s “Mind Children”, mentioned a nearly identical non-learning wasp scenario, except it was a tarantula wasp.

You might also find the series: How does my brain work interesting.

I especially like the last video on Growing evidence of brain plasticity.

Merlin - Jul 30, 2013:

You might also find the series: How does my brain work interesting.

I did. I started into the various videos you mentioned. I thought each one had at least one item of value, and the first one of the brain series one stayed on my mind quite a while as thought-provoking, which is a good sign.

()
Sebastian Seung: I am my connectome
http://www.ted.com/talks/sebastian_seung.html
I didn’t know the state of the art—that the only species that has had its connectome figured out so far is C. elegans (= Caenorhabditis elegans, a nematode).
()
Henry Markram: A brain in a supercomputer
http://www.ted.com/talks/henry_markram_supercomputing_the_brain_s_secrets.html
Interesting about the moon-horizon phenomenon, and the operation of anaesthetics. I hope that when researchers finally try to locate “the rose” inside the brain by examining axon activity that they try something a *lot* simpler, though, like a solid red, featureless wall!
()
1. Daniel Wolpert: The real reason for brains
http://www.ted.com/playlists/1/how_does_my_brain_work.html
I don’t know that I’d agree that the first and main (“real”) purpose of the brain is movement: memory might be a lot better function, even in an amoeba, since an amoeba would find it useful to remember previously encountered dangers even though it has no muscle-controlled limbs. The cup stacking video is incredible. The sea squirts example of sea squirts eating their own brain is cogent. This video brings up a good point, indirectly: saying that “the brain works like such-and-such” is awfully misleading: Wolpert’s direction is based on probability formulas and probably arrays or tensors, but even if part of the brain works that way, especially the cerebellum, pattern recognition probably does not work that way. That’s partly why I’ve become particularly interested in vision lately, and I believe that operation of the brain is based on very different principles than motion/arrays/sensors. Now I’m starting to realize in more depth why some people say that intelligence is a *collection* of abilities, each of which could be very different.

I’m trying to get time to look at Boris Sunik’s material, posted in another recent thread. Sorry, Boris, I’ve been short on time.

By the way, nobody seemed to catch a new potential problem I later noticed with my new proposed definition of “intelligence”: the wording should make clear that efficiency also applies to learning, not just to processing. So let’s try another update: “Intelligence with respect to a given goal is the ability to do the following things efficiently for attaining that goal: learning,  and processing real-world data.”

And more errata: the speed of electronic circuits is 10^6 faster than biological neurons, not 10^3. I knew that. :-(

While neurons work on the order of milliseconds, silicon oper-
ates on the order of nanoseconds (and is still getting faster).
That’s a million-fold difference, or six orders of magnitude.
The speed difference between organic and silicon-based minds
will be of great consequence. Intelligent machines will be able to
think as much as a million times faster than the human brain.
(“On Intelligence”, Jeff Hawkins with Sandra Blakeslee, 2004, page 223)