This is the NeuroDimension Newsletter, which you are receiving because you requested to stay informed about new developments at NeuroDimension. If you would like to stop receiving these newsletters, please see the bottom of this newsletter for instructions.
In this issue...
|
Happy Holidays Sale - Save 20%
NeuroDimension is excited to offer the Happy Holidays Sale! From now through December 30th, all NeuroDimension software can be purchased at 20% off the regular list prices.
This is an excellent time to purchase NeuroSolutions, TradingSolutions, a NeuroSolutions or TradingSolutions add-on, or one of our genetic products!
|
|
But don't wait too long - this offer ends December 30, 2005. Use offer code 53554 when placing your order. You can place your order now using our online order entry system at http://www.neurosolutions.com/order/
Note: This offer cannot be used in conjunction with any other special offer or on Bowfort Technologies' Fuzzy Candlesticks add-on for TradingSolutions.
|
Neural Network Course Complete!
We successfully completed another neural network and NeuroSolutions course in early November. We always enjoy getting to know our customers better, especially when it gives us a chance to discuss their applications. Once again, we received excellent reviews from our attendees and have never had anyone tell us they would not recommend our course to others.
|
|
The next Neural Network course is being planned for Spring 2006. If you would like more information on our courses, please visit: http://www.neurosolutions.com/products/course/
|
Looking for a partner for your application?
|
Have a new application in mind or an existing one that could be enhanced with neural networks or genetic optimization, but not quite sure how to utilize these powerful technologies?
|
|
NeuroDimension is looking for partners with exciting applications that need a hand in bringing those applications to the market. We have over 14 years of experience in the field of artificial intelligence and have worked on many different applications of this technology. If you are interested, please send us a brief description of your application along with your contact information to info@nd.com and we will have someone contact you about a potential partnership.
If you have any other questions, please do not hesitate to contact us at info@nd.com.
|
|
NeuroSolutions Tip Box: Leave-N-Out Training
One of the most common problems people encounter when developing a neural network model is that they do not have enough data to
adequately train the network. A rule of thumb we often use is that the training samples to weights ratio (i.e., the number of
training rows divided by the total number of weights in the network) should be at least 5, but preferably 10 or higher.
|
|
|
For example, if your neural network is a one-hidden-layer MLP with 4 inputs, 6 hidden processing elements and 2 outputs, you should
have at least 180 training rows (5 * (4*6 + 6*2) = 180). What if you only have 200 rows total? Once you pull out 20% for cross
validation and 20% for testing that only leaves you with 120 training rows. One option would be to reduce the size of the network
by reducing the number of inputs and/or hidden PEs. Another approach would be to use the Leave-N-Out algorithm, which is now
included as part of the Train module of NeuroSolutions for Excel.
The Leave-N-Out algorithm trains the network multiple times, each time omitting a different subset of the data and using that
subset for testing. This enables you to use all of your data for training and all of your data for testing. The testing results
are still out-of-sample since the rows being tested are not used to update the network weights during that particular training
run.
The documentation for Leave-N-Out has been updated with the 5.02 release of NeuroSolutions for Excel. It contains more of the
details of how the algorithm works and what is reported. One thing to be aware of is that the cross validation set is not being
used the same way as in the standard training routine. The cross validation set is actually the test set which is shifted between
the training runs. The fewer rows you tag as cross validation, the more training runs will be required and the longer it will take
to run. However, the testing results may be better with a smaller number of rows being tested since each run will have more
training data.
One other thing to consider with Leave-N-Out is the weights used for the testing are those obtained at the end of the training run
(and not those with the lowest cross validation error), so it is important not to set the number of epochs too high as this may
cause the network weights to be overtrained before each test. You may want to perform a standard training run first to see how
many epochs the network trains before the cross validation set reaches the lowest error and use this number for the Number of
Epochs setting within the Leave-N-Out panel.
|
|
We appreciate your feedback! Please send us your comments or suggestions concerning this newsletter, our web site, or part of the NeuroDimension product line. Write to us at: newsletters@nd.com.
Have questions about NeuroDimension products? Send your questions to: info@nd.com
This issue and previous issues of this newsletter are available on the NeuroSolutions web site at: http://www.neurosolutions.com/newsletters.html
If you would prefer not to receive these newsletters or subsequent product updates from NeuroDimension, please reply to this letter with the subject heading changed to the word REMOVE.
You can also contact us with questions or removal requests via mail at:
NeuroDimension, Inc., 3701 NW 40th Terrace, Suite 1, Gainesville, FL 32606
Thank you again for your support of NeuroDimension products!
|
|
