wgi@aplcomm.jhuapl.edu (William G. Innanen) (04/09/91)
Recently I posted a request for opinions on the NN package "NeuralWorks Professional II/PLUS" for the Macintosh and other machines. I received several resopnses, unfortunately none of them from Mac users. The responses (and the resulting correspondence) were nonetheless useful. Following this is the original posting, most of the responses, and some follow up correspondence. I've learned from NeuralWare that there is a current 50% (!) discount for qualifying academic customers. This is only for NeuralWorks, not for the other packages such as DesignerPack. Thanks again to all who answered. And if there are _any_ Mac users of NeuralWorks out there, I'd still like to hear from you! Bill Innanen wgi@aplcomm.jhuapl.edu =========================== Original Posting =========================== I am beginning the process of proposing a project involving the use of neural nets. I would like to hear from anyone who has experience with the package "NeuralWorks Professional II/PLUS" by NeuralWare, Inc. of Pittsburgh. I am planning on using it with a Macintosh IIfx so I would be most interested in Mac users' opinions. Since the package is available for several other machines, comments from other users would also be appreciated. If you have some other favorite neural net package for the Mac, I would be interested in hearing about that, also. Since I'm not a regular reader of this group, please respond by e-mail. I will post a summary if response (and interest) warrant. ========================== Replies ===================================== From: hdavids@mswe.dnet.ms.philips.nl Subject: Re: NeuralWare Bill, On your request for user experiences with NeuralWare: we have used the Neuralware package for some time now. Overall, our impression was that it is a good starting point if you are for the first time going to experiment with NNs. Lots of options let you get started quickly, but once you are into serious work, it is usually better to write your own software so you know exactly what's going on. We have used the accompanying Designerpack to transfer network sw from the PC to our Suns. The PC (a 386 with floating point chip) proved to be too slow for real-life problems. The generated sources that you have to compile on the Sun are very hairy - not something you'll love to update. Henk ------------------------------------------------------------------------ From: William G. Innanen <wgi@aplcomm.jhuapl.edu> To: hdavids@mswe.dnet.ms.philips.nl Subject: Re: NeuralWare Henk, Thanks for the reply! Your comments on the DesignerPack were especially welcome. I hadn't planned on ordering it immediately, relying instead on the "Flash Code" feature that comes bundled with NW/Pro. Any comments? I'm planning to use the code to develop Mac applications. I'll be building my demo programs on a Mac IIfx (68030 @ 40 MHz) with 8 Mb memory and an accelertated graphic board. Do you think that will be fast enough to keep me from growing a beard while I watch it? Bill Innanen ------------------------------------------------------------------------ From: hdavids@mswe.dnet.ms.philips.nl To: wgi@aplcomm.jhuapl.edu Subject: Re: NeuralWare Bill, the Flash Code feature is something that I don't know about. Guess I'll have to read the release notes more carefully that came with the latest update. The performance of the Mac should be sufficient for moderate-size problems, although a floating point co-processor would help a lot. Is that installed? We tried to teach a network to recognize hart contours from X-ray images, and for that type of problem you need all the cpu power you can get. To give you an idea: we used a fast Sparc server, with lots of memory, but it still took around 2 days (!) to get to the point where the network would stabilize. Now this may not exactly be the easiest type of application, but one lesson we learned from this is that you shouldn't jump too hastily to the conclusion that the learning process has reached a stable state, or that your network does not converge. The trouble in general with NN's is that you want to try various combinations of networks and training algorithms, adjusting the parameters a little on many runs. Without dedicated hardware this takes a long time, sometimes to long to be practical. However, if you're doing something that has not been tried before, you have no option. On a more optimistic note: I also tried to teach a standard backprop net to predict the results of the major soccer league here in Holland. That network had 36 input nodes, 24 in a hidden layer and 2 output nodes. Using the outcome of the 188 games played sofar, the training took 45 minutes on the Sparc. And the Sparc is able to compute the exp() function in 4 micro- seconds... Good luck, Henk ------------------------------------------------------------------------ To: hdavids@mswe.dnet.ms.philips.nl From: wgi@aplcomm.jhuapl.edu Subject: Re: NeuralWare Henk, Yes, a Mac IIfx has a floating point co-processor (all Mac II's do). The fax I got from NeuralWare, Inc. gives the following features for the Mac version of NeuralWorks Professional II/Plus (I'm still waiting for the literature via snail). Are they different from what you have? * A graphical interface with an icon based tool palette. * "FlashCode(tm)" which instantly converts your fully trained networks into highly portable "C" code routines. * "ExplainNet(tm)" ... * "BackProp Builder(tm)" ... * "Neural Probe(tm)"... * Support for 20 nn paradigms through "InstaNet(tm)" * Support for nn chip development. * Flexible, expandable architecture. ... * Over 900 pages of docs... * On-line help... Thanks for all of your help. Its nice to know that someone has gone on before me. Bill Innanen PS: From the look of things they gave the Trade Mark Office a "WorkOut(tm)" :-) WGI ------------------------------------------------------------------------ From: hdavids@mswe1.dnet.ms.philips.nl To: wgi@aplcomm.jhuapl.edu Subject: NeuralWare and FlashCode Bill, according to the docs that came with our latest update we now also have the NeuralWorks Professional II/Plus software, and that includes the FlashCode command. So, that puts an end to my bewilderment. I haven't tried it yet (have to find a machine to convert the floppies first), but when I do I will let you know how the code looks. According to the documentation however, FlashCode can only be used to create code for *fully trained* networks. Designer Pack is able to give you code that you can run on an other machine to actually train a net. So, in order to use a faster cpu to shorten the training process you still need Designer Pack. That would not be so bad if only Designer Pack weren't so expensive: about a year ago, when we bought the Neuralworks software, the Professional II package for IBM-PC cost $1495, but Designer Pack was $1995. I still think that it was money well spent, but the ratio between the the two strikes me as a bit odd. Probably something to do with their pricing policy - still being able to sell the SUN version for a nice sum. Can you please change whatever you intend to post with these findings in mind? Thanks, I'll be in touch later. Henk ======================================================================== From: schultz@halley.serc.3m.com (John C. Schultz) Subject: Need Opinions on NeuralWorks Professional II/PLUS We have been using NeuralWare Nworks II for about a year and are quite pleased with both the software, the performance, and the support personnel. We run it on a sun 3 which would be a little slower than a MAC IIFX I would think. You will probably want to use multi-finder on the MAC since training for several hours is not unusual. Running heavy duty, compute bound programs in the background may provide unacceptable interactive performance on a MAC. I would highly recommend the one week training class even though it is rather expensive ($1500). I like the diagnostic tools provided with Nworks. The software is also reasonably priced (lots cheaper the Hecht-Nielson for example). The only concern I would have with using a MAC (other than the multi-tasking issue above), is that NeuralWare does their developement on a SUN, then ports to the PC and then to the MAC. This may delay your software revisions. ======================================================================== From: s72026@zeus.usq.EDU.AU (gaynier ronald) Subject: Re: Need Opinions on NeuralWorks Professional II/PLUS I have a copy of NeuralWorks Pro II, it is a fine if you're not interested in doing what isn't already being done. In other words, if you're interested in using neural networks as a tool in your project it may suit your needs, however if you are interested in doing neural network research for its own sake you may be disappointed. On a IBM compatable who really need extended memory and a math co-processor, its a large program and doesn't seem real efficient. That may be because it tries to do too many things. It is a good teaching tool however and does allow you to create and try networks very easily. ======================================================================== William G. Innanen Johns Hopkins Univ/Applied Physics Lab (Internet) wgi@aplcomm.jhuapl.edu