NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/comp/ai/neuraln/2964

Text File | 1992-07-26 | 3.8 KB | 76 lines

Newsgroups: comp.ai.neural-nets
Path: sparky!uunet!zaphod.mps.ohio-state.edu!caen!destroyer!ubc-cs!unixg.ubc.ca!kakwa.ucs.ualberta.ca!alberta!arms
From: [e-mail] (Bill Armstrong)
Subject: Re: Neural Nets and Brains
Message-ID: <arms.712184394@spedden>
Sender: [e-mail] (News Administrator)
Nntp-Posting-Host: spedden.cs.ualberta.ca
Organization: University of Alberta, Edmonton, Canada
References: <arms.711935064@spedden>> <[e-mail]>     <arms.711986585@spedden> <[e-mail]>     <arms.712073050@spedden> <[e-mail]>
Date: Sun, 26 Jul 1992 20:59:54 GMT
Lines: 63

[e-mail] (Bill Skaggs) writes:

>[e-mail] (Bill Armstrong) writes:

>   >Aside from brain modelling using purely boolean signals, there are
>   >other potential uses of ALNs un studying the brain; namely
>   >explaining what is happening in nerve cell collections.  If the
>   >axons have action potentials, then it would still be possible,
>   >even if there are continuous signals at other places in the
>   >system, to study the relationships among axon firings using
>   >boolean methods.  One would have to relate the firing of axon A to
>   >firings of others B, C, ... at previous time instants.  In this
>   >way, one would factor out the continuous parts of the operations
>   >and concentrate on the functional modelling just based on boolean
>   >operations and time delays.
>   >
>   >Does anyone know if this has been tried?

>Hmm, well.  It's at least been *thought of*.  The problem is that when
>there are potentially thousands of inputs to a cell, and the cell
>integrates inputs over a hundred or so time steps, it takes way too
>much space to represent the operations in Boolean terms.  If you're
>trying to get the relationships biologically correct, the only
>reasonably efficient way to represent neural operations is in terms of
>difference equations.

I wouln't automatically assume that ALNs couldn't do it.  For example
Allen Supynuk used ALNs with 1740 input bits in predicting future
dynamic states of a robot.  I admit that the integration over 100 time
steps would give you 100000 boolean inputs, which is probably at or
beyond the limit of feasibility of our current software.  We have
trained about 100000 nodes though.

I think there are lots of things that can simplify the apparent
complexity.  The refractory period of an axon suggests that there may
be little activity on an input most of the time.  One could thus take
the latest signal at a synapse together with the time since it
occurred. (The actual signal can be summarized in this way even if it
is analog provided the actual form of the signal is fixed.)  You could
also take prior signals to the most recent ones too.  The above ideas
have been tried out quite successfully in our work with Dr. R. B.
Stein to develop designs for prosthesis controllers for spinal cord
damaged patients.

One promising related result, which has just been presented at the
First International Conference on Functional Electrical Stimulation in
Sendai, Japan, by Stein et al shows how the EMG (muscular) activity in
a cat's leg could be predicted from the signals in one of the cat's
sensory nerves using ALNs.  (The hope would be to do away with the
pressure pads in the shoes of patients, and use their internal
signalling, but that's still a way off.)

If there is anyone out there in netland who is interested in trying to
discover the relationships among neurons using ALNs, I would be
pleased to try to help in designing the experiments from an ALN point
of view.

Bill
--
***************************************************
Prof. William W. Armstrong, Computing Science Dept.
University of Alberta; Edmonton, Alberta, Canada T6G 2H1
[e-mail] Tel[phone] FAX [phone]