Hi Bruce,

On Fri, 29 Dec 2000, Bruce S. O. Adams wrote:

> be the common folklore. How much information does it take to represent a
> single neuron sufficiently accurately to model a brain? There are some very
> detailed modelling efforts which I think go further than necessary.

You are probably right that many efforts go further than necessary. But
that is far from saying that the most simplistic model is adequate. The
more I look at it, the more it is apparent that a single neuron, with its
dendritic tree and synaptic sites is quite a computational beast. You're
not just looking at a single operational unit there. The dendritic tree
allows for quite a number of interesting interactions between responses
at various sites, both forward and backward. Abstractions are possible,
but I would be very surprised if a single biological neuron could be
reliably abstracted into a single abstract neural network node in the
A.I. sense. The chemical factories with their frequency-tuned inhibitory
or excitatory exchanges and balances, acting on different time-frames, at
each synaptic receptor site, and in `compartments' of the dendrites and
the neuron are so specialized for their tasks that it is obvious that
they perform non-trivial work in the local information processing tasks.
Thus you would have to use an abstract neural network to model the
behaviour of a single biological neuron. At that point, the effort to do
the computations in the neural network probably start to outweigh the
efforts of a slightly more complex, but tailor-made (hence efficient)
computational model of the neuron.

I personally don't believe that every minor electrical fluctuation and
arbitrary positioning of components needs to be emulated in precise
detail. But the specialized functions of types of receptors, their
number, their average locations, and interaction functions should
probably be represented by some functions to be included in a model. It
is also quite important that the time-component be included, i.e. that
event timing is significant. Traditional neural networks tend to convert
everything to batches and represent response frequencies as an amplitude.
That is now known to be insufficient - you really need event timing,
since single responses (too few to compute a rate representation from)
interact during mental processes. The book "Spikes: Exploring the Neural
Code" by Fred Rieke et al. was dedicated to the purpose of waking up the
neuroscience community to this previously much-overlooked fact.

In my opinion, the minimal implementation of a neuron that can achieve a
reliable emulation of the significant components of the information
processing in a biological neuron is a specialized spiking neuron, with a
number of intrinsic functions, both for the response and modulatory
functions, combined with connectivity components that implement a number
of specialized functions for their transfer functions, local modulations
at different time-scales, interaction with neighbouring connection sites,
and interaction with the neuron's soma. That's pretty much what I listed
in the early draft I posted to minduploading.org. The exact number of
simultaneous functions is something I don't know yet, and it probably
varies per neuron.

The functions themselves can be reasonable representations of more
detailed operations that are either modeled from their observed responses
or from known kinetics. There will have to be probabilistic components.
Perhaps some of the functions can be safely simplified to a degree that
makes them more pallatable to digital computers, but possibly they will
have to be refined with more details instead. It is quite possible that
many simultaneous nonlinear dynamic (perhaps discontinuous) functions
need to be solved during the emulation of a single neuron. It is very
useful to find efficient ways to do this with the goal of emulating the
temporal occurrance of relevant information processing events with
sufficient precision to satisfy the needs expressed in the
above-mentioned book and elsewhere. One such efficient way is the
event-predictive emulation I've been working on, and which I mentioned in
the draft. As a result I was less interesting in counting neurons, than I
was in counting the number of prediction computations per neuron, and the
number of instructions per prediction, to come up with the computational
demands for a software implementation.

> Perhaps the Qualitative Reasoning Neuron is the way to go. I'm currently
> hedging my bets on simple artificial neural network models. I'd be
> interested in
> any evidence or suggestions for experiment which would rule out either
extreme.

I'd say the simple extreme has already been ruled out by neuroscientific
evidence - although it is still probably possible to represent a single
biological neuron by a network of artificial ones. More work needs to be
done to reign in the needs for complexity, especially hard core
information theoretic investigations, the only way to determine the actual
significance that particular features of the biological functions can have.

> Your model differs from mine. Perhaps you could provide a comparison
> in terms
> of number of neurons and the memory and processing required to simulate each?

You should get a pretty good idea of those by reading the draft again as
I improve it. I'll post to MURG again when I upload a new version of the
draft.

> I got the impression you were aiming for more accurate modelling than I
> was. I also
> note on your page you are using floating point arithmetic which even in
> these modern
> times of superscaler processors still costs us a lot in terms of performance.

You're right. There are definitely many efficiency gains as the code is
optimized, including decisions regarding the matching of computational
operations to the performance of the chosen hardware.

> I see most of these as essential research tools rather than as the kind of
> abstractions
> we can scale up to cover an entire brain. Hence my continuance with the
> 'old hat' of more simplistic neural networks of proven capability but
> massively scaled in terms

I've tried to explain my point of view above. I don't think that an
emulation that more faithfully achieves the computational power inherent
in biological neurons is beyond current feasibilities at all.
Demonstrating feasibility is in fact the point of my draft-in-progress.

> of size and complexity. Can you cite something which points to their
> insufficiency at doing
> the job?

I hope I did, starting with the book by Fred Rieke et al. There is a lot
more, but that requires some digging, which I will do if you need me to.

Cheers,
Randal

_______________________________________________________________________
RANDAL A. KOENE
Neural Modeling Lab, Department of Psychology - McGill University
randalk@..., www.psych.mcgill.ca/perpg/stds/rk/
minduploading.org, Lab:(514)-398-4319, Home:(514)-767-6406
_______________________________________________________________________