marwood@ncs.dnd.ca (Gordon Marwood) (05/12/91)
I should be interested in any opinions anyone may have on Neuralworks Professional II Plus running under MS-DOS. A colleague of mine is considering the acquisition of this software to assist in studies of association of positional data, e.g. association of new data with multiple existing tracks of previous data. My apologies if I am covering old ground, but I have only just started to look at comp.ai.neural-nets. Gordon Marwood Internet: marwood@ncs.dnd.ca P.S. If any replies are mailed, please copy to bridgew@ncs.dnd.ca
kcj@goanna.cs.rmit.oz.au (Kate Juliff) (05/12/91)
In article <1991May11.232140.24961@ncs.dnd.ca> marwood@ncs.dnd.ca (Gordon Marwood) writes: >I should be interested in any opinions anyone may have on Neuralworks >Professional II Plus running under MS-DOS. A colleague of mine is >considering the acquisition of this software to assist in studies of >association of positional data, e.g. association of new data with >multiple existing tracks of previous data. I bought it and it is of no use to what I am doing at the moment. Its implementation of a Hopfield and Tank net is not standard and the description of such a net in the manual contains information which I found misleading. I have not looked in detail at the other NN models, but from a brief overview, I would say that the product is best used as an educational tool, that is using it to illustrate the different models, rather than for real applications. -------------------------------------------------------------------------- Kate Juliff VICTORIA UNIVERSITY OF TECHNOLOGY Email: kcj@matilda.vut.edu.au P.O. 64 FOOTSCRAY 3011 Fax: +61 3 689 4069 AUSTRALIA Phone: +61 3 688 4490 Disclaimer: The opinions expressed are my own. ________________________________________________________________________
schultz@halley.est.3m.com (John C. Schultz) (05/13/91)
In article <5690@goanna.cs.rmit.oz.au> kcj@goanna.cs.rmit.oz.au (Kate Juliff) writes: In article <1991May11.232140.24961@ncs.dnd.ca> marwood@ncs.dnd.ca (Gordon Marwood) writes: >I should be interested in any opinions anyone may have on Neuralworks >Professional II Plus running under MS-DOS. A colleague of mine is >considering the acquisition of this software to assist in studies of >association of positional data, e.g. association of new data with >multiple existing tracks of previous data. I bought it and it is of no use to what I am doing at the moment. Its implementation of a Hopfield and Tank net is not standard and the description of such a net in the manual contains information which I found misleading. I have not looked in detail at the other NN models, but from a brief overview, I would say that the product is best used as an educational tool, that is using it to illustrate the different models, rather than for real applications. I disagree. I think that the NeuralWare product is most useful for applications rather than research. Previous comments that I have read in this newsgroup indicate that researchers are somewhat constrained by the limited flexibility of the underlying network model in NeuralWare. (Or they just prefer to roll their own :-) For my part, I use back propagation networks on applications (prediction, modeling) and NeuralWare works quite well. I particularly like the diagnostic tools, the ability to deduce the important input paramters, and the ability to generate run-time `C' for an on-line implementation of the network. I also think that NeuralWare has some very good personnel on their staff. I did attend the 1 week training class which I found very useful in being able to use Nworks quickly and efficiently. I can't specifically address the Hopfield and Tank network issue raised above but NeuralWare's back propagation (with options of many of the various enhancements in the technical literature) works quite well for me. I use both a Sun 3 and a MSDOS version. -- John C. Schultz EMAIL: schultz@halley.serc.3m.com 3M Company, Building 518-01-1 WRK: +1 (612) 733-4047 1865 Woodlane Drive, Dock 4, Woodbury, MN 55125 How to include the taste of Glendronach in a multi-media system?
wouter@duttnph.tudelft.nl (Wouter Schmidt) (05/14/91)
At our research group we bought Neuralworks Professional II Plus running on a Sun 4 platform. We agree that this software is usefull as long as you use the standard buildin network paradigms to solve your application. It is visually a very good tool for demonstrations and may serve as a good intro- duction environment to neural networks. However as a pure research tool it is (for our group) of no use. Try for example to implement the back propagation rule for radial basis functions or to implement a complete new network paradigm within neural works. Our experience is that most of your coding effort is wasted to understand and to implement the neural works requirement. The documentation on this subject is poor and if your 'new' idea does not work you do not know if you made a programming error or your 'new' idea is of no use. A succesfull implementation of a new network idea depends on a number of software layers in neural works and modification must be done in all of these layers. An understanding of your adaptions to the software may not completely be clear due to the complex software structure. As argued before in this newsgroup if you are doing research in developing new neural networks techniques you will often end up coding your own neural networks environment in a general purpose language like C or C++. This argument still holds for neural works professional ! Neural works does not support batch processing. A workstation is occupied as long as your simulation takes because the software requires a graphics window open. In our group we must share workstations and long simulations (which is quite often the case in the neural networks learning) prevent other users to login and use their own environment. Furthermore a three year license and supports costs us about US $14.000. The latest update we got does not seem to be much more than a change in the user interface, which is not much value for money for this expensive software license. I (we) agree with the previous posted statement: > I have not looked in detail at the other NN models, but from a brief > overview, I would say that the product is best used as an educational > tool, that is using it to illustrate the different models, rather than > for real applications. -- +--------------------------------------------------------------------------+ | Wouter Schmidt FI-PH | | | Pattern Recognition Group, | Room : 229 | | Faculty of Applied Physics, | Phone : 31-(0)15-78 46 94 |
sci240s@monu6.cc.monash.edu.au (mr w.j. ho) (05/14/91)
wouter@duttnph.tudelft.nl (Wouter Schmidt) writes: >However as a pure research tool it is (for our group) of no use. Try for >example to implement the back propagation rule for radial basis functions or >to implement a complete new network paradigm within neural works. Our >experience is that most of your coding effort is wasted to understand and to >implement the neural works requirement. The documentation on this subject is >poor and if your 'new' idea does not work you do not know if you made a >programming error or your 'new' idea is of no use. >A succesfull implementation of a new network idea depends on a number of >software layers in neural works and modification must be done in all of these >layers. An understanding of your adaptions to the software may not completely >be clear due to the complex software structure. >As argued before in this newsgroup if you are doing research in developing new >neural networks techniques you will often end up coding your own neural >networks environment in a general purpose language like C or C++. This argument >still holds for neural works professional ! ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ What about using a high level language for the description and definition of neural network architectures ? 2 examples of commercially available packages are : 1) AXON by HNC Inc; and 2) ANSPEC by SAIC. Anyone has experience with these or any other similar languages or tools? Personally, for reasonable speed and flexibility to do some real experiment and work, I would prefer a hardware based neuro-coprocessor. For example, the Balboa or Anza from HNC, the Delta II from SAIC, and many other companies offering similar boards. Get whichever one that suits your needs, budget, and etc ( company support, add-ons, libraries, etc etc ). -- ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ ^ Wey Jing Ho Tel: 61-3-5732567 E-mail : sci240s@monu6.cc.monash.edu.au ^ ^ Physics Dept., Monash University ( Caulfield Campus ), Melbourne, AUSTRALIA ^ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^