![[Image of Me]](Res/Other/DJR01bL.jpg)
|
Under Construction |
|
|
|
|
For now I'll start some links to helpful resources... Mainly they will be links that I personally might find useful, or links to my own contributions. The hope is that some of what I find useful may also be useful to others, but that is not a driving value for the time being. For now this page is here mainly to get MY Neural Network information and thoughts organized.
Just a warning: Much of my understanding behind this organization framework stands in direct 180-degree oposition to what has been related --in what would be considered-- mainstream neural network "papers and publications" for the last 15 - 20 years.
So: For now (November-2007 and the immediately future) everything here is "coming soon..." There is a temporary to-do list further down on this page.
Robotics Links:
coming soon...
Neural Network Links:
NETLAB version 0.1 is an environment for building and running neural networks. It gives you 700 "implants", 8000 connections, and a handful of input/output functions called "prostheses". It allows you to connect anything in this environment to anything else. The "implants" can be configured as "neurons" which have special inputs for training weights within their synapse-space. Like everything else in the environment, anything can be connected to these weight training inputs. Implants configured as "neurons" have many configurable parameters such as attack, decay, and response type (constant, Amacrine), and these parameters are configurable on an individual basis. NETLAB is basically a "hardware" model produced by a "hardware hacker", and is in many ways geared to the hardware hacker. NETLAB is NOT a collection of mathematical models. If you're looking for a package that will let you run the "black-box" models that are around [...] this is NOT the package for you. NETLAB v. 0.1 comes with a special runtime for Back-propagation (BAP.EXE), but this is a last minute throw-in, and is the only "black-box" model included with NETLAB [v 0.1].When I uploaded this first version of Netlab in 1990, I included permission for the file to be distributed for non-revenue generating uses. By today's standards this early version has many disadvantages. For example, the neuron output values and weight values are only 8-bits wide (very little resolution). Still, the package introduced early ideas about temporality, and the ability to use stimuli (axon levels) to produce error levels for weight-training. That is, you can connect an axon to a training input of a neuron to act as positive and negative reinforcement stimuli for the neuron. In this early document version this was referred to as "affective feedback" [sic].
The package has remained continuously available on the various pre-internet networks and their Internet archival counterparts. The following is a listing of some locations where you can find this distribution on the web. This early version should be considered purely for the historical perspective it provides to an upcoming Netlab release. It should not be thought of as a good means of producing neural networks. That is, it is included here purely for its historic significance relative to a new release of Netlab that will be made within the next year or so.
For some reason Google has stopped indexing web file locations holding this early version (I am quite clueless regarding the hows and whys of search engines). Following is a list of some locations where the original Netlab version 0.1 programs and documents can be downloaded. These all contain the same content, but they are made available in different compression formats: