Greetings
from NeuroDimension!
The World Leader in Neural Network Software
This issue of the NeuroDimension
newsletter highlights the latest developments at NeuroDimension, as well as
providing insights on how and when to use neural networks.
In this issue you’ll find:
What’s New
and News?
* Neural Network Course
Discount
* Newsletter Index Added to
Website
Note: You are
receiving this newsletter because you requested to stay informed concerning new
developments at NeuroDimension. If you would like to stop receiving these
newsletters, please see the bottom of this newsletter for removal instructions.
=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=
The next neural network course is scheduled for October 21 -
October 25, 2002 at the Grosvenor Resort, located in the Walt Disney World
Resort in Orlando, Florida. A 10% discount is available for early registration,
so be sure to register today!
Our course format allows both novice and advanced users to
find a suitable course. Offered courses include: "Introduction to
NeuroSolutions", "Fundamentals of Neural Networks and
NeuroSolutions", and "Advanced NeuroSolutions". The courses
include a copy of our interactive book, Neural and Adaptive Systems:
Fundamentals Through Simulations.
For details on this new offering, or to sign-up from the
Internet, see http://www.neurosolutions.com/products/course/oct_2002.html
For general ND course information, see http://www.neurosolutions.com/products/course/
For more information and samples of the interactive book,
see http://www.neurosolutions.com/products/nsbook/
Trying to find more information about a particular topic in
neural networks or NeuroSolutions? Check out the new Newsletter Index added to
our recently redesigned NeuroSolutions website. All of our previous
NeuroDimension newsletters topics are now listed by volume for easy reference.
Plus, a special topic index has been added to quickly find every issue that
addresses individual topics.
Visit the new newsletter index today at http://www.neurosolutions.com/newsletters.html
=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=
Designing Neural Networks
Neural
networks are very powerful tools when used properly. However, they are only as
powerful as the data they are given to work with. In fact, the most common
reason that people encounter problems applying neural networks is that they do
not have enough data or they do not take the time to preprocess that data
properly.
Limited
Data Availability
Neural networks are powerful because they are nonlinear and non-parametric, but
this flexibility requires enough data to properly train the system (the more
flexibility in the system, the more data required to “restrict” the system to
accurately map the data).
The
relationship between the number of features per data sample and the number of
data samples (e.g. rows versus columns in Excel) gives a first order
approximation to whether you have sufficient data for a neural network model. A
good rule of thumb is that you should have a minimum of 10-20 times more rows
than columns.
Therefore,
image processing, Fourier spectrum analysis, and other high dimensionality
tasks may require significant preprocessing to reduce the input size if you
don’t have a large database of records.
Symbolic
Data Expansion
Another
situation where limited data can come into play is symbolic processing such as
natural language recognition. Symbolic inputs create situations where the input
data cannot be adequately represented in a numeric format without symbolic
expansion.
Symbolic
expansion converts a single column with distinct values into unique columns for
each possible value. For example, if a column can contain 100 different words,
symbolic expansion would convert this column into 100 different columns, each
with a true or false value indicating whether that value has occurred.
It is important
to recognize when symbolic expansion may be needed. Just because data is
numeric doesn’t mean that data contains numerical rather than symbolic
relationships. For example, one column of data could be voting districts
numbered 1 through 6. While this data is represented as a number, it should
still be considered to be symbolic. This is because there is no numerical
relationship between the individual numbers. In other words, the relationship
between districts 1 and 2 is typically not the same as that between 2 and 3 or
3 and 4.
Once it is
understood when symbolic expansion is needed, this can affect how much data is
needed to effectively train a neural network. Using the rule of thumb described
above, 200 rows of data may be enough to train a neural network based on 10
data columns. However, if one of those columns is symbolic and contains 21
different values, symbolic expansion would turn this into a total of 30 data
columns, making it difficult to effectively train this network.
Handling
Limited Data
This is not to say that neural networks cannot and have not been applied
in these areas when limited data is available, just that typically these types
of problems need to be broken down into subtasks. Neural networks should be
applied to only those subtasks that make sense. For example, in image
processing, you can preprocess the image to pull out key features such as
center of mass of the object, boundary characteristics, etc. and these features
can be applied to the neural network. For symbolic processing, sometimes the
task can be redefined as determining when only a few values are present, rather
than any value.
Even when large amounts of data are available, it is always
worth considering whether the problem is being represented in the best possible
way. Before “throwing data into a neural network to see what comes out”,
consider specifically what you are trying to do. Are you trying to predict the
likelihood of certain conditions or looking for any condition? What
characteristics in the input data would you look at if you were trying to solve
this same problem by hand? These types of questions can be very useful in
modifying the inputs to get the best possible results from a neural network.
=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=
Customer Spotlight
NeuroDimension, Inc. has recently completed the first phase
of a project aimed at extracting the fetal electrocardiogram (FECG) and other
important FECG features from the electrocardiogram of a pregnant woman (MECG).
Using the customizable DLL features of NeuroSolutions,
NeuroDimension has implemented a recently introduced blind source separation
(BSS) algorithm named Mermaid, which uses information theory learning to
separate mixed independent sources such as fetal and maternal ECGs. Based on
our results, the algorithm achieved very good performance in outperforming
other BSS competitors, as well as the Multiple Reference Adaptive Noise
Cancellation (MRANC) algorithm traditionally used in the separation of FECG and
MECG signals. These results have been validated with both synthetically mixed
data (for which we have a measure of demixing) and real ECG data (for which we
have developed a quantitative performance measure) from 32 subjects.
NeuroDimension has also developed front-end signal
processing hardware to enhance the signal-to-noise ratio (SNR) of conventional
ECG monitors to deliver a better quality signal and improve the success of FECG
separation.
=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=
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: feedback@nd.com
Have questions about NeuroDimension products or training
services? Send your questions to: info@nd.com
Some mail readers may display an attachment at the bottom of
this newsletter. This is typically caused when the mail reader is set to
display text only. The original, formatted version is sometimes included as an
attachment.
This issue and previous issues of this newsletter are
available on the NeuroDimension 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.
Thank you again for your support of NeuroDimension products!