| View previous topic :: View next topic |
| Author |
Message |
Greg Heath
Guest
|
 Posted: Sat Aug 16, 2008 3:18 pm Post subject: Re: help with testing and validating data |
 |
|
On Aug 16, 6:32 am, chan <chantelojinn...@yahoo.co.uk> wrote:
| Quote: | I have tried the above code and it doesn't work.
|
Of course it doesn't work. It wasn't meant to be a plug-in
solution to your homework problem.
It is a template to guide you.
Try to understand it and you will see that
1. the doc and help commands should be removed and used
before running the script in order to understand newff.
2. there is no assignment of H which, in general, is
chosen by trial and error using a validation set to
estimate error.
3. there is no loading of validation and test sets.
Try again and if you have trouble understanding error
messages, we can help.
| Quote: | I need to train, test
and validate data using tansig and logsig.
|
You mean train and validate; then test. Validation is part
of the design process and is used to determine adequate
parameter settings such as number of hidden nodes and number
of training epochs.
If you don't need to determine these, you don't need a
validation set.
| Quote: | I can train the data from
the source dat1.txt using tansig and logsig using the code from my
first post, I just dont know how to test and validate the data. I am
completey new to matlab and really need help.
|
I am traveling and don't have acssess to MATLAB. Therefore I
cannot help unless you provide more details.
If you have a trained net and validation data Pval,Tval
yval = sim(net,Pval);
eval = Tval-yval;
MSEval = mse(eval)
can be used to determine H by trial and error. Once you
are satisfied with H, use the test data, Ptst,Tst
ytst = sim(net,Ptst);
etst = Ttst-ytst;
MSEtst = mse(etst)
to estimate generalization error.
Hope this helps.
Greg
P.S. have you seen my pretraining advice post? |
|
| |
|
 Back to top |
Greg Heath
Guest
|
| Posted: Sat Aug 16, 2008 3:20 pm Post subject: Re: help with testing and validating data |
|
|
On Aug 16, 5:04 am, chan <chantelojinn...@yahoo.co.uk> wrote:
| Quote: | Thanks for your reply. Im new to the groups and do not know how to
crosspost to multiple groups.
|
You just did. Look above to the line
Newsgoups: comp.soft-sys.matlab, comp.ai.neural-nets
Hope this helps.
Greg |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Sat Aug 16, 2008 3:52 pm Post subject: Re: About train Neural Network function |
|
|
On Aug 15, 12:59 pm, "Manos " <ezoul...@in.gr> wrote:
| Quote: | Dear Greg,
Thank you for your valuable help. I try to apply your
suggestion and what I understod from my web search.
I type the following code to Matlab
%Some explanation
%This make a special split that I need. variables including
%Result are the features and with YResults are the Targets
%This part works fine as I wish but I give it here for
%consistence
% I have 1639 cases with 12 features for network inputs and
%1 feature for network target
[HeadLines,NMIMF,MSIMF,MSIMF1,MSIMF0,NMSIMF1,NMSIMF0,MSIM1,MSIMY1,MSIM0,MSIMY0,NMSIM1,NMSIMY1,NMSIM0,NMSIMY0,NMTRF,MTRF,MTRF1,MTRF0,NMTRF1,NMTRF0,MTR1,MTRY1,MTR0,MTRY0,NMTR1,NMTRY1,NMTR0,NMTRY0,NMVALF,MVALF,MVALF1,MVALF0,NMVALF1,NMVALF0,MVAL1,MVALY1,MVAL0,MVALY0,NMVAL1,
NMVALY1, NMVAL0,
NMVALY0,MT1,MTY1,MT0,MTY0,NMT1,NMTY1,NMT0,NMTY0,ResultsNM,ResultsM,ResultsNMold,ResultsMEid1,ResultsMEid0,ResultsNMEid1,ResultsNMEid0,YResultsMEid1,YResultsMEid0,YResultsNMEid1,YResultsNMEid0,YResultsNM,YResultsM,YResultsNMold,TotalSimResult,YResultsSimTotal,TotalTrainResult,YResultsTrainTotal,TotalValResult,YResultsValTotal,TotalTestResult,YResultsTestTotal,YResultsTrainTotalNum,YResultsSimTotalNum,YResultsTestTotalNum,YResultsValTotalNum]
split_to_nn_sets_and_sim_for_suspicious_spec(textdata,data,0.05,0.6,0.2,1);
%This creates the network using the Train part only
%TotalTrainResult matrix of 934x12
%YResultsTrainTotalNum matrix of 934x1
net11 > newff(TotalTrainResult,YResultsTrainTotalNum,10,{},'traingd');
|
doc newff
help newff
Note that inputs and outputs must have the same number of columns.
Therefore, transpose your matrices.
Hope this helps.
Greg
P.S. You removed the crosspost to comp.ai.neural-nets. I
replaced it. |
|
| |
|
| Back to top |
chan
Guest
|
| Posted: Sat Aug 16, 2008 4:12 pm Post subject: Re: help with testing and validating data |
|
|
| Thankyou so much for your help, I really appreciate it. |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Mon Aug 18, 2008 7:14 pm Post subject: Re: overshoot in RBF |
|
|
On Aug 18, 2:36 pm, Greg Heath <he...@alumni.brown.edu> wrote:
| Quote: | On Aug 12, 4:26 pm, Prime Mover <eple...@hotmail.com> wrote:
Can anyone cite a reference with discussions or explain why overshoot
often happens in classification when using Radial Basis FunctionsNeuralNetworks? I have run several models with the same training set
and results are maximum values are always beyond 1 (often 1.15, 1.20).
This is because the default output activation units are PURELIN
and there is no provision for an alternayive.
These values turn the classification results poor as to the point of
discredit the models?
No. Classification results should be reliable as long as
the next highest output is not close to the highest value.
With {0,1} targets, the classification assignment is
made to the class with the largest posterior probability
estimate, regardless of the value.
If you wish the posterior estimates to obey
0 < P(i|x) < 1
and, for mutually exclusive classes,
SUM(i = 1:c){ P(i|x) } = 1
it would be useful to have the options to use
the LOGSIG or SOFTMAX activation functions.
However, MATLAB does not provide those options.
An alternative, beside modifying source code, is
to use the default PURELIN for training but use
LOGSIG or SOFTMAX post-training for classifying
nontraining data.
|
so that posterior probability estimates can be
used to quantify the confidence of the classification.
Hope this helps.
Greg |
|
| |
|
| Back to top |
Gavin Cawley
Guest
|
| Posted: Mon Aug 18, 2008 7:31 pm Post subject: Re: overshoot in RBF |
|
|
On 18 Aug, 20:14, Greg Heath <he...@alumni.brown.edu> wrote:
| Quote: | On Aug 18, 2:36 pm, Greg Heath <he...@alumni.brown.edu> wrote:
On Aug 12, 4:26 pm, Prime Mover <eple...@hotmail.com> wrote:
Can anyone cite a reference with discussions or explain why overshoot
often happens in classification when using Radial Basis FunctionsNeuralNetworks? I have run several models with the same training set
and results are maximum values are always beyond 1 (often 1.15, 1.20).
This is because the default output activation units are PURELIN
and there is no provision for an alternayive.
These values turn the classification results poor as to the point of
discredit the models?
No. Classification results should be reliable as long as
the next highest output is not close to the highest value.
With {0,1} targets, the classification assignment is
made to the class with the largest posterior probability
estimate, regardless of the value.
If you wish the posterior estimates to obey
0 < P(i|x) < 1
and, for mutually exclusive classes,
SUM(i = 1:c){ P(i|x) } = 1
it would be useful to have the options to use
the LOGSIG or SOFTMAX activation functions.
However, MATLAB does not provide those options.
An alternative, beside modifying source code, is
to use the default PURELIN for training but use
LOGSIG or SOFTMAX post-training for classifying
nontraining data.
so that posterior probability estimates can be
used to quantify the confidence of the classification.
|
I expect the NETLAB toolbox for MATLAB implements LOGSIG and SOFTMAX
for RBF networks given that Prof. Nabney wrote a very clear paper on
the subject
http://www.ncrg.aston.ac.uk/netlab/
NETLAB is a jolly good bit of kit, I used to use it a lot until I
wrote my own, and it is even free! |
|
| |
|
| Back to top |
ronarld.chan@googlemail.c
Guest
|
| Posted: Mon Aug 25, 2008 9:39 pm Post subject: Re: rate of convergence of RBF on solvling one dimesnsion he |
|
|
Hi guys.
Does anyone know about rate of convergence of RBF on solvling one
dimesnsion heat equation with a smooth initial input functon? Thanks!
Cheers
Ron |
|
| |
|
| Back to top |
Phil Sherrod
Guest
|
| Posted: Tue Aug 26, 2008 5:33 pm Post subject: Re: rate of convergence of RBF on solvling one dimesnsion he |
|
|
On 25-Aug-2008, "ronarld.chan@googlemail.com" <ronarld.chan@googlemail.com>
wrote:
| Quote: | Does anyone know about rate of convergence of RBF on solvling one
dimesnsion heat equation with a smooth initial input functon? Thanks!
|
The answer is 43. I'll define my units later. ;)
Seriously, there are numerous ways to implement RBF networks. Also, the speed
would depend on the roughness of the surface being modeled. So there is no way
to answer your question with the information you've provided. If you want to
post some data, maybe one of us can run some tests.
--
Phil Sherrod
(PhilSherrod 'at' comcast.net)
http://www.dtreg.com (Decision trees, Neural networks, SVM and Genetic
modeling)
http://www.nlreg.com (Nonlinear Regression) |
|
| |
|
| Back to top |
ronarld.chan@googlemail.c
Guest
|
| Posted: Mon Sep 01, 2008 10:59 am Post subject: Re: rate of convergence of RBF on solvling one dimesnsion he |
|
|
On 26 Aug, 13:33, "Phil Sherrod" <PhilSher...@REMOVETHIScomcast.net>
wrote:
| Quote: | On 25-Aug-2008, "ronarld.c...@googlemail.com" <ronarld.c...@googlemail.com
wrote:
Does anyone know about rate of convergence of RBF on solvling one
dimesnsionheatequation with a smooth initial input functon? Thanks!
The answer is 43. I'll define my units later. ;)
Seriously, there are numerous ways to implement RBF networks. Also, the speed
would depend on the roughness of the surface being modeled. So there is no way
to answer your question with the information you've provided. If you want to
post some data, maybe one of us can run some tests.
--
Phil Sherrod
(PhilSherrod 'at' comcast.net)http://www.dtreg.com (Decision trees, Neural networks, SVM and Genetic
modeling)http://www.nlreg.com (Nonlinear Regression)
|
Hey Phil,
Thanks for your advice about my topic. My approach is following E.J.
Kansa's suggestion, (http://rbf-pde.uah.edu/kansaweb.html). Basically,
his idea is to tranform PDE, like heat equation to a system of ODEs
and then solve this stiff ODEs by using any implicit schemes. Suppose
we provide an initial conditon is smooth function, Can I still have
any kind of rate of convergence?
Cheers
Ron |
|
| |
|
| Back to top |
ronarld.chan@googlemail.c
Guest
|
| Posted: Mon Sep 01, 2008 11:07 am Post subject: Re: About adaptive radial basis function on solving Time-dep |
|
|
Dear guys,
So far on the literature on solving Time-dependent Burger Equation by
adaptive radial basis function, I only found two papers about it, one
is by Scott Sara and another one is by Driscoll. However, in their
papers they don't show the convergence of their methods. I am
wondering if there is any work about adaptive radial basis on Time-
dependent Burger Equation by adaptive radial basis functions. Also,
any theories about this kind of ideas?
Cheers
Ron |
|
| |
|
| Back to top |
Guest
|
| Posted: Fri Sep 19, 2008 8:54 pm Post subject: Re: Spreading the word |
|
|
On Sep 19, 9:12 am, Jean-François Michaud <come...@comcast.net> wrote:
| Quote: | On Sep 18, 7:19 am, "roger.l...@gmail.com" <roger.l...@gmail.com
wrote:
On Sep 11, 9:34 pm, "Robert Miller" <rsmil...@compmore.net> wrote:
An article on Forth by Stephen Pelc appeared today on Dr. Dobbs website.
Great article!
I thought the article was interesting but it seemed that the flow was
a bit off-kilter [...]
Jean-Francois Michaud
|
The article at http://www.ddj.com/embedded/210600604 was
interesting and valuable enough for me to link to it from
http://mentifex.virtualentity.com/fp080917.html --
the MindForth Programming Journal for 17.SEP.2008.
In that journal entry citing Mr. Stephen Pelc, I
record the careful elaboration in Win32Forth of
what is one of MindForth's most intricate
cognitive tricks -- the ability to comprehend
a fact-finding question and answer it from the
AI knowledge base with one simple word, if
warranted by the KB: an affirmative "YES"!
http://mentifex.virtualentity.com/mind4th.html
has a new "kbSearch" module that searches
the knowledge base to answer "YES" to
a question or to ignore it otherwise.
http://mentifex.virtualentity.com/exhibit.html
is guidelines for putting MindForth on display
as artificial intelligence for science museums.
Arthur |
|
| |
|
| Back to top |
bleh
Guest
|
| Posted: Sat Sep 20, 2008 9:29 pm Post subject: Re: Spreading the word |
|
|
mentifex@myuw.net wrote:
| Quote: | http://mentifex.virtualentity.com/exhibit.html
is guidelines for putting MindForth on display
as artificial intelligence for science museums.
|
Step 1: Write poster about AI winter
Step 2: Cite mentifex and MindForth
Step 3: Profit
LOLZ. |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Sat Oct 11, 2008 7:49 pm Post subject: Re: Neural Networks Character Recognition |
|
|
On Oct 11, 11:31 am, "kobe " <cyber_k...@hotmail.com> wrote:
| Quote: | Hey Guys,
I was wondering if anyone knew how I could obtain information/code regarding handwritten character recogntion using the Neural Network toolbox, I've run the programs on Matlab's file exchange but they're not really helpful. Has anyone done a similar project? I have 2 recognise digits 0-10 and train the network. I've read the various feedforward algorithms such as Back Propagation but how do I implement them for character recognition?
Thanks.
|
I've forwarded this to comp.ai.neural-nets
Hope this helps.
Greg |
|
| |
|
| Back to top |
z.zz123@live.com
Guest
|
| Posted: Sat Oct 25, 2008 6:58 am Post subject: Re: Neural Networks Character Recognition |
|
|
On Oct 11, 10:49 pm, Greg Heath <he...@alumni.brown.edu> wrote:
| Quote: | On Oct 11, 11:31 am, "kobe " <cyber_k...@hotmail.com> wrote:
Hey Guys,
I was wondering if anyone knew how I could obtain information/code regarding handwritten character recogntion using the Neural Network toolbox, I've run the programs on Matlab's file exchange but they're not really helpful. Has anyone done a similar project? I have 2 recognise digits 0-10 and train the network. I've read the various feedforward algorithms such as Back Propagation but how do I implement them for character recognition?
Thanks.
I've forwarded this to comp.ai.neural-nets
Hope this helps.
Greg
|
You may have a look at papers of Geoffrey E. Hinton
http://www.cs.utoronto.ca/~hinton/
and Yann Le Cun http://yann.lecun.com/ |
|
| |
|
| Back to top |
Isaac
Guest
|
| Posted: Sun Nov 16, 2008 5:35 pm Post subject: Re: A Critique of Prof. Hubert Dreyfus' "Why Heideggerian AI |
|
|
"Publius" <m.publius@nospam.comcast.net> wrote in message
news:Xns9B57F131CB7B0mpubliusnospamcomcas@69.16.185.250...
| Quote: | "Isaac" <groups@sonic.net> wrote in
news:491f9f87$0$33506$742ec2ed@news.sonic.net:
Minsky, unaware of Heidegger's critique, was convinced that
representing a few million facts about objects including their
functions, would solve what had come to be called the commonsense
knowledge problem. It seemed to me, however, that the deep problem
wasn't storing millions of facts; it was knowing which facts were
relevant in any given situation. One version of this relevance
problem was called "the frame problem." If the computer is running a
representation of the current state of the world and something in the
world changes, how does the program determine which of its represented
facts can be assumed to have stayed the same, and which would have to
be updated?
Dreyfus is pointing out one consequence of the lack of a useful definition
of "intelligence."
Actually, he is doing much more than that in his paper. I just posted a |
portion of its background section, but his paper sets forth what he believes
is why AI fails and how he (and certain philosophers/researchers he relies
on) thinks intelligence works. I will email you the paper for your
reference. Let me know if you are interested in me posting my critiques of
his paper for your response. Maybe you will defend his positions?
| Quote: | It is problem which plagues most programs for producing
AI (which is not to deny that much progress has been made in that
endeavor).
We may define "intelligence" as, "The capacity of a system to generate
solutions to novel problems," and "problems" as, "Obstacles or impediments
preventing the system from attaining a goal."
I defy you to contrive a definition of Intelligence that works. For |
example, using your current definition above, the Earth would be intelligent
because it is a system with the capacity to generate solutions (e.g.,
extremely complex, yet stable atmospheric weather, ocean currents, etc.) to
solve novel problems of, for example, maintaining a stable global
temperature in the face of many (thousands) changing (novel) variables that
are constant obstacles preventing the Earth (Gia?) from attaining her goal
of minimizing temperature differences globally. There are many similar
examples that use your language but are not considered to be intelligent to
anyone reasonable in science. Care to update your definition or defend it?
| Quote: |
Introducing goals into the definition gives us a handle on the "frame
problem": the problem is framed by the current goal.
|
Goals are really related to the frame problem in that the "frame" that
matters is the one that reflect "reality" in the context of your priorities,
experience, and world model (e.g., a filter) as a situated agent. Goals are
just one priority, but goals to not realy drive perception, they mostly seek
to manipulate the frame to achieve a desired result. Loosly, I think the
frame problem is much more about constructing a useful and tractible model
that the situated agent can use towards building a plan to achieve its
goals.
| Quote: | Attention is paid only
to world states which bear on the system's goals (as a background
process).
|
Of course, goals to play an important role in how to focus attention, and to
some extent this colors the frame problem, but I do not see how it drives it
exclusively as you put it. If I look at a movie I have no goals other than
to be entertained, however, I clearly create the proper frames needed to
understand and appreciate the movie and its meaning. In that case, I am
actually trying to discover the goals of the movie, and not my own, to
understand it. No goals on my part, but the frame problem seriously exists.
| Quote: | If not enough information is in hand to solve the current problem, then
the
system returns to "the world" to gather additional information. (There is
no need to "store millions of facts." Facts are gathered as they are
needed, i.e., in light of the present goal and problem).
|
I don't think anyone would say that classic AI would not return to the world
to gather more facts to add to its "millions of facts". The issue that
Dreyfus says is the problem with AI is that it creates rules that are
representations (or symbols) and are compartimentalized, both of which he
says the Philosopher Heidegger espouses, which Dreyfus and his set of
philosophers/researcher say is not the case. I think every Intelligent
system will end up effectively having a constantly evolving set of millions
of "rules", so that is not the question. Do you have any counter examples?
Cheers!
Ariel B.- |
|
| |
|
| Back to top |
|
