Q.
How much do I have to know about artificial intelligence to use the AI
Trilogy?
A. Absolutely nothing, because the software itself requires very
little knowledge, and we’ll teach you what little you need to know. Our
neural networks aren’t like the ones from the 1980’s that required in-depth
knowledge and lot’s of experience-based “tweaking” to get the model right.
Although many companies still sell the neural networks using the algorithm known
as “backpropagation”, we abandoned it years ago except in our legacy product
NeuroShell 2.
Q.
Why don’t you have any white papers on the site about AI Trilogy
applications?
A. The AI Trilogy is so simple to use that almost all of our users
build their own applications, receiving only free advice from us. We do not
make hundreds of thousands dollars selling highly priced “complete
solutions” using flashy salespeople that show up in your office. Therefore,
since customers save money collecting the example data and building the
models themselves, we have never had to write any white papers.
Q.
Can I distribute models built in the AI Trilogy in my own software?
A. Neural network models can be distributed anywhere using the NeuroShell
Run-Time Server, which is included in the AI Trilogy package. It lets you
call your neural network models via dynamic link library (DLL) calls from your
programs, or by calling COM objects. You can also distribute your models
inside of a spreadsheet.
Our genetic algorithm
program GeneHunter does not have any run-time because each use is usually
separate and within a spreadsheet. However, if you do program our genetic
algorithm via calls to DLLs, you may distribute such applications only
within your company unless you purchase a custom distribution license.
Q.
Aren’t neural networks black boxes?
A. Neural networks make models that use huge, complicated mathematical
equations that are virtually impossible to reverse engineer or understand.
In this sense, you could call them black boxes. But then so must multiple
regression, especially non-linear regression, be a black box. Although
linear regression builds simpler polynomial equations, many stat books take
several chapters to explain how to interpret these simple linear
polynomials. The truth is people use software every day that they do not
understand the inner workings of. Neural networks are in automobiles today, and
you drive those cars, even though you probably understand neither the neural
networks in them, nor the details of the internal combustion engines that power
them.
Q.
What kinds of neural networks does the AI Trilogy use?
A. We have two types. The first type, which we call the “neural
method” uses our super-fast TurboProp2 neural network. TurboProp2 builds
great “global” models of the training data.
The second type, which we
call the “genetic method” uses neural networks developed by Dr. Don Specht.
In the NeuroShell Predictor, we use Specht’s General Regression Neural Nets
(GRNN). In the NeuroShell Classifier we use Specht’s Probabilistic Neural
Nets (PNN). Our GRNN and PNN neural networks are trained by genetic algorithm so that
they are excellent at finding the contribution of each variable in the
training data, and eliminating worthless variables. Although much slower
training that the “neural method”, the “genetic method” builds great “local”
models of the training data.
Q.
Don’t neural networks take a long time to train?
A. The older “backpropagation” training method was indeed very slow.
Generally, our “neural method” trains in between 2 and 60 seconds, depending
on the number of training examples and independent variables (inputs). The
“genetic method” is much slower because it is more local, and training speed
decreases arithmetically as more training data is used. It may take several
hours if there is a large amount of training data.
Q.
Don’t neural networks tend to “curve fit” the data?
A. We prefer to use the term “over-fit” because almost any time you
build predictive models from previous history you are curve-fitting.
Regression curve fits too. The problem comes when you make your model so
non-linear that you incorporate small variations of the data (usually called
noise) into the models (we call that “over-fitting”). So neural networks, like
all non-linear modeling, can indeed over-fit if you aren’t careful (a good
part of the free advice that we provide helps users avoid over-fitting).
However, the neural method has built-in techniques to help reduce
over-fitting. The genetic method is even better, because it does a “hold one
out jackknife” even while it is training. That means that if there are N
examples in the training data, curves are only being fit through N-1
examples in order to evaluate the one held out. This means that the genetic
method is often robust on new “out-of-sample” data, even if it seems less
effective on the training data.
Q.
Why won’t you supply a demo that I can test on my own data?
A. We don’t do much advertising of our software, because given the
relative low prices we charge for each unit, expensive advertising doesn’t
make sense. We prefer to get sales prospects by word of mouth from satisfied
customers. It is therefore paramount that everyone who uses the software on
their own problems be as satisfied with the results as possible, because we
cannot afford bad word of mouth. Most people think they can just run their
data “as is” through a non-linear model and the results speak for
themselves. Nothing could be further from the truth. In order to get good
results, most real world problems need some expert data formulation or
preprocessing, and a good deal of care regarding how to choose the
variables, the training sets, and the test sets. We provide this “modeling”
expertise free for purchasers, but given our low profit margin, we cannot
afford to provide it for everyone who wants a free trial, many of whom
aren’t serious software buyers anyway. The consulting and modeling expertise
that we provide is an integral part of our software price, and we do not
want to unbundle it, because we would rather lose a sale than sell to
someone who is likely to be unsuccessful because they do not want to take
advantage of our expertise. We have been selling neural networks full time this
way since 1988, and that is why we have out-lasted dozens of competitors
that existed in 1988. Even IBM tried to compete against us years ago and
finally abandoned selling general purpose neural network software.
Q.
How will I know if the AI Trilogy will solve my problem?
A. You should talk to us about your problem – it’s free, and we
encourage it. If our software won’t solve your problem, we’ll tell you.
Otherwise we’ll tell you exactly what you need, and the steps you’ll need to
build the application. Look, we’ve done this hundreds of times, and if you
can explain your problem, we can quickly see how our AI Trilogy will solve
it. Believe it or not, we actually HATE to sell the AI Trilogy to people who
won’t discuss their applications first, unless they just want it for
educational purposes.
Q.
When would I want to purchase the legacy NeuroShell 2 product?
A. We sell this mostly to teaching professors, computer science
students, or anyone else who is more interested in “playing” with neural
networks than in solving real problems. It has 16 neural network paradigms and
backpropagation architectures, and allows access to inner components. As
such, it is great for learning about neural network technology, but far more
difficult to use for real problems. We are not going to build any future
releases of NeuroShell 2, as it is essentially a ‘frozen” legacy product.