Im very much fascinated by biological Neural Networks. But because Im
a computer scientist, and not a biological or medical, I dont know a
great deal of the biological aspects of Neural Networks. Ive made a very
simplified requirement specification of a computer model for Neural
Networks.
I would like to ask someone to take a look at it (form a biological point
of view), and tell me what else I need to know, or whats wrong in it.
Any other comments are welcome
Thanks, Ray
.
.
.
.
N N EEEE U U RRR AA L N N EEEE TTT W W OO RRR K K SSSS
NN N E U U R R A A L NN N E T W W O O R R K K S
N N N EEEE U U RRR AAAA L N N N EEEE T W W O O RRR KK SSS
N NN E U U R R A A L N NN E T W W W O O R R K K S
N N EEEE UUUU R R A A LLL N N EEEE T W W OO R R K K SSSS
Computational Neural Network Model
Requirements Specification
(C) 1998 By R. Storm
1. Neural Networks Overview
---------------------------
1.1 Biological Neural Networks
------------------------------
1.1.1 Information about biological neurons
------------------------------------------
- Each neuron acts on its own.
- A neuron collects signals at it's synapses by summing all excitatory
and inhibitory influences upon it.
If the excitatory influences are dominant the neuron sends a message
to other neurons. This is decided by the neuron threshold function.
(e.g. step, ramp, sigmoid or gaussian)
- The neuron value will degrade through time, otherwise threshold will
always be exceeded.
- An axon carries information through series of action potentials.
- A synapse represents the junction between an axon and a dendrite.
- Synapses are made during in the early life of a organism. Probably
when that neuron is stimulated enough connections are made with nearby
neurons. These connections are not reversible.
1.1.2 Information about biological neural networks
--------------------------------------------------
- Over 100 billion (10^11) neurons in a human brain.
- Maximal 1000 synapses on the in & output of a human neuron.
- Parallel processing.
- Has a way of preventing information overloading or loss.
- All neurons are positioned in a 3d space.
- Sensors are simply attached to the dendrites of a group of neurons.
- Actuators are attached to the axon of a group of neurons.
1.2 Simulation of Biological Type Neural Networks
-------------------------------------------------
To successfully design a simulated model of a biological neural network
it's necessary to discuss the way of implementation of al the in paragraph
1.1 mentioned properties:
1.2.1 Computational Model of a biological neuron
------------------------------------------------
- Each neuron acts on its own.
This is a big problem if the model is designed for a computer with a
single CPU. So there must be a timed method of the Neural Network class
which can perform the actions for each neuron. This can be done in several
ways, the best way will probably be:
Each neuron that has been stimulated sets a flag, which indicates it
needs to be updated. Then the method that updates the network processes
in a standard pattern all neurons.
- A neuron collects signals at it's synapses by summing all excitatory
and inhibitory influences upon it.
If the excitatory influences are dominant the neuron sends a message
to other neurons. This is decided by the neuron function. (e.g. step,
ramp, sigmoid or gaussian)
When a neuron threshold value is exceeded, it uses it's numbered list with
synapses containing pointers to next neurons. When updating a triggered
neuron send a signal to all other neurons to which it is connected.
- The neuron value will degrade through time, otherwise threshold will
always be exceeded.
Two datamembers of the Neuron class will keep track of the reduction of
the neuron value. The write-off delay will indicate the number of update-
cycles there are before reduction begins. Another keeps track of the
number of cycles that has passed since the last update. The power of
reduction is determined via the reduction function, which can very much
look like the neuron threshold function.
- An axon carries information through series of action potentials.
These series can and probably will occur between several update-cycles,
when the neuron is triggered multiple times.
- A synapse represents the junction between an axon and a dendrite.
A synapse is represented either by a class or just by a list or array
of dendrites with pointers to connected neurons.
- Synapses are made during in the early life of a organism. Probably
when that neuron is stimulated enough connections are made with nearby
neurons. These connections are not reversible.
When a neuron is stimulated enough, it may create a synapse with the
most nearby neuron (possibly the one with the least activity).
This could be made reversible. The synapse is removed by a neuron, which
wasn't stimulated enough, and so it removes all synapses, using a
datamember that keeps track of it's activity.
1.2.2 Computational Model of biological neural networks
-------------------------------------------------------
- Over 100 billion (10^11) neurons in a human brain.
This is possible, but will use lots of memory an CPU-speed, even though
a computer is faster. Because the computer use the advantage of parallel
processing.
- Maximal 1000 synapses on the in & output of a human neuron.
This can be limited by a Neuron datamember, which represents the maximum
number of synapses.
- Parallel processing.
See paragraph 1.2.1, section: "Each neuron acts on its own".
- Has a way of preventing information overloading or loss.
The way of how information overloading is handled is described in para-
graph 1.2.1, section "The neuron value will degrade through time,
otherwise threshold will always be exceeded.".
For what concerns information loss (or forgetting). Humans have short
and long term memory. It probably has something to do with the degeneration
of the synapse connections (caused by aging of the organism). and neuron
thresholds (with their write-off delay). The long and short term memories
can be simulated by gives groups of neurons different neuron value
write-off
delays.
- All neurons are positioned in a 3d space.
The neurons will be ordered in memory (through arrays or lists) in
a box-like representation, of which the length, height, depth are
variable.
- Sensors are simply attached to the dendrites of a group of neurons.
Add a physical attribute to the main application (organism) to connect
input organs (keyboard, bitmap) to the neural network. The neural network
must have means (read: a method) to connect the sensor to the neural
network.
- Actuators are attached to the axon of a group of neurons.
Add a physical attribute to the main application (organism) to connect
output organs (screen, sound, printer) to the neural network. The neural
network must have means (read: a method) to connect the actuator to the
neural network.
---=== End ===---