Aaron Sloman
Last updated: 28 Dec 2006
Slightly modified and expanded.

In December 2006 I was asked to comment on the work of Jeff Hawkins whose Redwood Neuroscience Institute is

devoted to studying and promoting biologically accurate and mathematically well-founded models of memory and cognition.
and who has recently founded a company Numenta to develop 'a new type of computer memory system modeled after the human neocortex'. With science writer Sandra Blakeslee he wrote a book On Intelligence. I confess the book is still on my (very long) 'to be read' list, though I have read and heard quite a lot about it.

Although I am not sufficiently knowledgable about Hawkins' work to be able to comment with authority, I thought I would expose my impressions of the work, and its limitations, in the hope that if I have made mistakes they will soon be corrected.

This is a slightly expanded version of the reply I sent to the questions I was asked.

> I am doing a story on (Palm and Treo inventor) Jeff Hawkins' new startup
> Numenta, which is developing a new computer architecture based on his
> theory of how the brain works
> Are you familiar with his approach or know someone who might be who
> could comment on its viability?

I have skimmed some stuff of his, and listened to part of on online
presentation, but I don't know his ideas well.

He is obviously very bright, and not a charlatan. Just a bit ignorant
(like many other people who write about brains, intelligence, computers,
and the future of AI -- and I freely confess I have my own areas of

He seems to me to be an example of a familiar pattern in the history of
AI and neuroscience: someone who believes that if only you get the
mechanisms right by finding out how they work in nature, or by using
some general theoretical principle, then you'll be able to use those
mechanisms to produce a human-like machine.

As far as I can tell, like all the others with that kind of belief he
has failed to do what any good designer really should do, namely
identify precisely what the requirements are. That failure is
commonplace because people think they know what the requirements are,
and they think the requirements are obvious because we experience human
like intelligence at first hand all the time.

Despite all that, I don't think anyone has come close to producing a
complete and detailed set of requirements even for human vision, because
the full variety of functions of vision is not yet understood, and we
are miles away from knowing what sorts of mechanisms are capable of
performing those functions. I've been writing about that for many years,
e.g. in this 1982 paper:
    Image interpretation: The way ahead? (PDF)

and more recently in these slides on the functions of vision:
     A (Possibly) New Theory of Vision (PDF)

[Added 19 Apr 2007:
    Perception of structure 2: Impossible Objects (PDF)
    A shorter presentation.

But you won't have time to read all those so I'll just mention a few

    Humans use vision for control of rapid fluent actions, for control
    of careful intricate actions, for finding out what actions are
    possible in a particular situation, for seeing empty spaces, for
    seeing what's going on in another person's mind (puzzlement,
    interest, boredom, fear, joy, sexual interest, etc.), for seeing how
    a piece of machinery works, for learning how to do a dance, for
    sight-reading music, for enjoying a painting or natural scene, for
    debugging computer programs, for recognising unstable structures,
    for doing mathematics, ... and many more. Moreover, the competences
    of a human visual system are capable of being extended in many
    different directions, apparently by adding new sub-architectures
    -- e.g. learning to read a new language using a different script.

Current artificial systems can barely do a tiny fraction of all that and
most people working in the field have no idea how much of the task they
are ignoring.

I could say the same sort of thing about other kinds of human
competence: language, motivation, learning, actions, social
interaction, ...

Although I am not familiar with the details of Hawkins' work I was
quickly convinced that he was unwittingly ignoring most of this
complexity, and that made him over-confident about what could be
achieved using his mechanisms.

> Even if you are not familiar with it, what would you say are the odds of
> someone developing a computer in the next few years that can truly learn
> from observation, that is not programmed but rather creates its own
> model of the world based on sensory inputs, and that can finally crack
> hard computer problems such as image and pattern recognition?

All of those (and more) are already being done in narrowly restricted
domains, and in some cases they outperform humans (or normal humans).

But putting all of those together with all the rich detail of the
competence of a two year old child (scaling out, rather than scaling up)
is far beyond what will be achieved in the next few decades -- partly
because the requirements are so hard to specify. I have some examples in
this poster presentation from AAAI'06

> Hawkins distances his approach from Artificial Intelligence, which he
> characterizes as using brute computing power and logic to make computers
> seem intelligent through their behavior.  Traditional AI, he contends,
> does not care much about how the brain actually works as long it
> produces the right behavior on the part of the computer.  Is this a fair
> characterization of the AI approach?

This completely ignores the fact that 'how something works' is not an
unambiguous referring expression. When someone asks how my computer
works he may be interested in the quantum mechanical details underlying
the functions of transistors that a physicist can explain (or maybe
cannot yet explain). He may be interested in the story an electronic
engineer can give of how the digital circuits work. He may be interested
in finding out what the low level virtual machine architecture is
(intel inspired, implemented by AMD with their extensions) -- this will
support the same programs as older generations of machines with
different low level mechanisms. He may be interested in what operating
system I use (linux), how the file system works, what the scheduling
strategies are, what privileges there are and what effects they have on
intruders. He may wish to know which applications I am running and what
their software architecture is, what forms of representation they use,
what algorithms they use and how they communicate. If the machine is
doing mathematics, e.g. proving theorems, he may want to know what
axioms and rules it is using, what mathematical syntax it uses, what
search strategies it uses.

Some of those descriptions may be quite unlike anything that goes on in
brains. Others may be very like what goes on in brains -- e.g. when
programmers try to get their programs to model their own ways of
thinking about a problem, e.g. solving equations, making plans,
searching for solutions, etc. Much of traditional AI, far from not
caring about how humans work was inspired by attempts to model how
humans work. (Even a turing machine is a result of Turing's attempt to
specify in the most general way what mathematical thinking is.)

Some people were more concerned with faithful modelling than others.
E.g. Newell and Simon tried hard, McCarthy and Minsky just wanted to
capture high level competences. I think both have moved closer to trying
to understand in detail what humans do, but not by working up from
theories about what brain cells do, which I think will lead only to dead
ends if not combined with top down requirements analysis.

Of course, many people who pay lip service to biological inspiration
design artificial neural nets that are nothing like real brains: they
ignore most of the complexity and diversity in brains, and most of the
architectural richness.

And so far, the competences of artificial neural nets, except in very
narrow applications, are way behind the problem solvers, theorem
provers, program checkers, planners, game players, etc. and the expert
systems that underly huge amounts of internet technology, even if they
are not called AI.

> Do you think it is actually possible today to create a machine that
> mimics the memory architecture of the brain (the neocortex, to be
> exact)?  And if it were possible, would that trump the results we
> currently get from AI techniques?

I doubt it. Not 'today', and not in the near future.

I don't know if it will ever be possible because I don't know what
exactly the neocortex or any other part of the brain does, and I suspect
nobody else does either, e.g. insofar as a lot of biological information
processing is heavily based on chemistry, perhaps in ways nobody has
discovered yet.

I don't expect researchers in the next few decades to make very much
progress in understanding the requirements for human like intelligence
because that requires them to stop playing with their favourite
mechanisms and instead take seriously the attempt to find out what can
be learnt from linguistics, philosophy, education, anthropology,
psychology (many aspects), psychiatry, studies of what many other
animals can and cannot do as well as AI and neuroscience.

I also don't know whether the mechanisms produced by billions of years
of evolution are the only ones that can meet all the information
processing requirements as well as the energy consumption, weight,
and size requirements. For all I know any other design would require
a human brain to be replaced by a machine bigger than our planet.

Kurzweil's predictions make assumptions about how brains do their
information processing. If much of it is molecular he could be way out.

But we may get answers in future by doing the required research across a
broad spectrum of disciplines, and making people in all those
disciplines communicate with and understand one another far better than
they do now.

Of course, the task may just be too much for human minds. Eg researchers
need the equivalent of degrees in at least three, and preferably six

I've written much about all these points and more, scattered over my web
site. One day I may try to organise it better!

I hope this is of some use, but I won't be surprised or offended if it
doesn't help. if you have to fit your article into a few hundred words
I can't help you: we are discussing the most complex machine on earth --
and maybe in a much bigger space.

best wishes.



Maintained by Aaron Sloman
School of Computer Science
The University of Birmingham