Neural networks. Questions from the experts. - page 20

 
lasso:
What data or results do you need to provide so that you can concretely identify what the problem is?

Probably for starters.

1) Structure of the network: number of layers, neurons, weights

2) Volume of training sample and number of epochs

3) relative error of the network at the end of training

4) Parameters for initialisation of weights - the form of distribution of values and their variance.


I looked back through the thread, I understand about 1 and 2.

 
alsu:

Probably to start with

1) Network structure: number of layers, neurons, weights

2) Volume of training sample and number of epochs

3) relative error of the network at the end of training

4) Parameters for initialization of weights - the form of distribution of values and their variance.


Skimmed back through the thread, about 1 and 2 I see.

on point 3, if I understand you correctly, in the attachment.

on point 4, I can not find anything in the manual, I will dig further, but I think that the distribution is uniform over a range of values, e.g., [-1;1]

 
lasso:

But not to drastically change the test results! Do you understand?

Here are the results of the runs on the test period of 1 month:

-9337

+5060

....


And I take it this is on the training period? FANN?
 
joo:
Use GA.


Well GA is no stranger to the problem of paralysis.

By the way, I looked at your library with interest. There wasn't a thread discussing it? Any thoughts and questions....

 
Figar0:


1) Well GA is no stranger to the problem of paralysis either.

2) By the way, I looked at your library with interest. There wasn't a thread discussing it? Any thoughts and questions....

1) It's not alien. But this problem is much less relevant compared to other methods of optimizing/training NS.

2) There was no discussion thread specifically on my algorithm. Answered some questions here.

 
Artificial Neural Networks and Genetic Algorithms short advanced training course
and genetic algorithms"

Applications for short-term advanced training courses "Artificial Neural Networks and Genetic Algorithms" are closed
Artificial Neural Networks and Genetic Algorithms",
conducted by the Department of Continuing Education of the M.V. Lomonosov Moscow State University
V.V. Lomonosov Moscow State University on the basis of the Moscow State University Nuclear Research Institute
Moscow State University. Completers of the courses receive a state certificate of advanced professional training.
state certificate of advanced training.
Trainings will be held twice a week in the evening from 19:00.
Classes start February 25, 2011.

To learn more about the course programme, get more information and
Please click here to apply for the course:
http://www.neuroproject.ru/kpk.php
 
lasso:

on point 3, if I understand you correctly, in the attachment.

for item 4, I can't find anything in the manual, I'll keep digging, but I think the distribution is uniform over a range of values, e.g. [-1;1].

Yep.

% correct - is this on the training sample or on the test sample?

And another question: don't you think that for a network classifier 1 input is somehow quite... not enough?

 
Figar0:

And I take it this is on the training period? FANN?


1. Yes, this is FANN.

2. No, these are the OOS results of the same NS trained under the same conditions, on the same OPs.

 
alsu:

Yep.

% correct - is that on the training sample or on the test sample?

And another question: don't you think that for a classifier network 1 input is a bit... not enough?


)) Thank you for your consideration.

1. % correct - it's on a test sample. In the context of this TS -- 57% is good, 60% is very good, 65% or more is excellent.

2. Why not enough? Enough is enough. If I can divide this data (with dimension=1) into classes by linear or visual methods, why can't I reproduce this consistently with NS?

...............

Now I tried in Statistics 6 to classify the submitted training examples (TS) with a probabilistic neural network (PNNS).

Empirically I selected the smoothing coefficient = 0.05.

Then retrained it repeatedly. The results are stable and do not change from training to training.

If this is true, then a new question arises, how to transfer VNS for use with FANN?

 
VladislavVG:

As for SVM:

This medod will always find one single dividing plane ....

Good luck ....

Vladislav, thanks for the suggested method.

Here is an excerpt from the description:

Часто в алгоритмах машинного обучения возникает необходимость классифицировать данные. 
Каждый объект данных представлен как вектор (точка) в p-мерном пространстве (последовательность p чисел). 
Каждая из этих точек принадлежит только одному из двух классов.

Is this a prerequisite for this method?

After all, in my OPs, the classes are heavily mixed up:

And the dimensionality of my OPs equal to 1, too, as I understand it doesn't work to the plus side:

Стоит отметить, что если исходное пространство имеет достаточно высокую размерность, то можно надеяться, 
что в нём выборка окажется линейно разделимой.


=====================================================

If you are already using this method, maybe you could try splitting my data?