Purpose of a layer

Traveler eightwings at hotmail.com
Wed Jun 5 14:54:59 EST 2002

[The parent article by Gary Frank was posted to several newsgroups
including comp.ai. David Kinny, the comp.ai moderator, apparently
rejected the post which means that it will not appear on the other
newsgroups. I am reposting my response without comp.ai in the header.]

In article <3cfd5454_4 at corp.newsgroups.com>, "Gary Frank"
<garyfrankNOSPAM at itol.com> wrote:

>The brain consists of layers of neurons.  The neurons in a layer are
>connected to neurons in one or more layers in particular ways.  What is the
>purpose of a layer?  Why is the brain constructed of layers?

As I explain on my site, at the very least you must have a sensor
layer, a signal separation layer, a coincidence detection layer, a
sequence detection layer, a command selection layer and an effector
layer. A layered architecture makes it easier to route signal pathways
and manage connectivity.

>  Why do
>artificial neural networks consist of so few layers, when the brain contains
>so many? 

My understanding is that the connectivity cross-section of humans and
animals is no more than 6 or 7 neurons. Not what I would consider a

>I've heard of fully connected networks, where all neurons are
>connected to all neurons in the next layer.

Fully connected networks are a legacy of GOFAI. Within the context of
emulating animal intelligence, they're worthless. As are ANNs in

>Has any thought gone into connecting artificial neurons in a different way?

People have been experimenting with spiking neural networks for years.

> What would guide the way that the neurons are connected?

In one word, timing. One of the things neurobiology has taught us is
that neurons generate discrete signals called spikes. All spikes look
alike. There is only one thing that distinguishes one spike from
another: its time of arrival. And forget about rate coding. It has
been shown that, given the temporal resolution of neurons (about 1
millisecond), rate coding is out of the question because it is way too
slow to account for observed reaction times in humans and animals. The
only thing left is the temporal relationships between signals: two
signals can only be concurrent or sequential.

P.S. I have a feeling you expected a reply from me as this is right
down my alley.

Temporal Intelligence:

More information about the Neur-sci mailing list