|
NN.MDB
The Database that thinks
© 2007 Ben Margolis
Glossary
Terms in neural networking vary. Some people
like to use biological words for everything, like neuron and synapse (I've even
seen "soma," "axon" and "dendrite"). While at the newsgroup comp.ai.nn they
don't even call them "neurons." Instead they call them "units." (Which means its
really not a "neural network" anymore, wouldn't it be a "Unitarian Network?")
I personally have a problem with the word
"network." It may have been accurate in the 50's for the hardwired perceptrons
they were building then, but today it means "a computer network" and as soon as
you say "neural network" to anyone outside the AI community, they they think you
mean multiple computers.
That's why I started using the term "Neural
Cluster" to describe the relatively small MLPs we can build in NN.MDB. (Actually
I usually use the phrase "simulated neural cluster" when talking to clients.) I
also prefer to call it a "run" instead of an "epoch" and "data" as opposed to
"input set."
|
NN.MDB Term |
Abriv. |
Term used at Newsgroup Comp.AI.NN |
Implementation in NN.MDB |
Details |
|
|
|
|
|
|
Neuron |
n |
Unit |
A record in the Neural Table (nn1Neu) |
Neural ID number (nID) its current value (charge),
its weight, previous delta, x,y coord. etc. |
|
Neural Cluster |
NN |
Neural Network |
A record in the Network Table (tblNets), or the
subset (query) of all neurons and synapse related to that record. |
Network ID (NetID), MLP design, names of ext. data
sources, training history |
|
Synapse |
Syn |
Weighted Connection |
A record in the Synaptic Table (nn1Syn). |
Two nIDs, a NetID, layer and the synaptic weight. |
|
Synaptic Layer |
SynLyr |
|
A subset of all the Synapse of the NN in the same
layer, with their connected Neurons. (qryCurrNNSynLyr) |
SynLyr1, contains Input and Hidden Neurons and
Synapses, SynLyr2, contains Hidden and Output Neurons and Synapses. |
|
FactData |
|
Training Set |
Any Access table or Query that contains training
facts, with expected output. |
|
|
TestData |
|
Verification Set |
Any Access table or editable Query that contains
training facts, with blank columns for NN output and with expected
output. |
(usually a subset REMOVED from the FactData) |
|
RunData |
|
|
Any Access table or editable Query that contains
runtime inputs with blank columns for NN output. |
Actual runtime data source. |
|
|
|
|
|
|
Run |
|
Epoch |
One full execution of any datasource through the NN |
|
|
|
|
|
|
|
network architecture |
|
|
The exact number of neurons in each layer of the NN |
I:2, H:10, O:1 would mean 2 input Neurons, 10
hidden, and 1 output. |
|
