Neural network performance problem
显示 更早的评论
I am doing a speech recognition project to classify 4 words; using MFFC.
Since speech divided into frames and features are extracted; often the number of frame is not constant; hence number of features also varies.
1)What is the best best neural network for such problem : Static or Dynamic (Time series)?
I have tried static network; since the input feature vector space has to be constant; i just convert the less feature vector to largest obtainable feature vector by padding zeros to the less feature vector.
2) Is it a good technique for such problem where the number of feature is not constant?
Now some question about the problem i am facing with static network:
1) "GO" and "STOP" words are training well; providing good validation result and test result.
2) "LEFT" and "RIGHT"; among these to words maximum times one of them is training well; in validation and test maximum time they are misclassified between them; I mean LEFT is misclassified as RIGHT and vice-versa.
3)Input matrix dimension is 1157x352 where 352 are the training example. Any comment. Using "nprtool" and hidden neuron is 10.
4)Almost all time validation is stopping the training; Is it a good sign??
5) what is best the training algorithm for such problem?
I want ur precious comment and experience share about such problems.
采纳的回答
Greg Heath
2012-11-17
0 个投票
Neural network performance problem
Asked by Daud on 16 Nov 2012 at 16:09
% I am doing a speech recognition project to classify 4 words; using MFFC.
% Since speech divided into frames and features are extracted; often the number of frame is not constant; hence number of features also varies.
% 1)What is the best best neural network for such problem : Static or Dynamic (Time series)?
Probably Static
% I have tried static network; since the input feature vector space has to be constant; i just convert the less feature vector to largest obtainable feature vector by padding zeros to the less feature vector.
% 2) Is it a good technique for such problem where the number of feature is not constant?
No. It is a bad technique. Try
1. Reducing your input dimensionality
2. Estimating missing values using the average, median or imputation.
Search the net using terms like 'missing values' or 'imputation'
Wikipedia might be a good source..
% Now some question about the problem i am facing with static network:
% 1) "GO" and "STOP" words are training well; providing good validation result and test result.
% 2) "LEFT" and "RIGHT"; among these to words maximum times one of them is training well; in validation and test maximum time they are misclassified between them; I mean LEFT is misclassified as RIGHT and vice-versa.
% 3)Input matrix dimension is 1157x352 where 352 are the training example. Any comment. Using "nprtool" and hidden neuron is 10.
Reduce the input dimension
% 4)Almost all time validation is stopping the training; Is it a good sign??
Validation stopping prevents the net from trying to memorize the idyosynchracies of the training data at the expense of performing well on nontraining data.
In other words: Use the training data to indirectly minimize the error on nontraining data by using the validation data to represent the important characteristics of nontraining data.
% 5) what is best the training algorithm for such problem?
It depends. Always try the MATLAB default first.
% I want ur precious comment and experience share about such problems.
Glad to help.
Thank you for formally accepting my answer.
Greg
4 个评论
Daud
2012-11-25
Greg Heath
2012-11-28
PCA for regression or curvefitting (NEWFIT OR FITNET)
PLS for pattern recognition or classification (NEWPR OR PATTERNNET)
Sometimes I start with STEPWISEFIT to get a preliminary feel for the data.
Greg Heath
2012-11-29
Sorry. Cannot help with this one.
Greg
更多回答（0 个）
类别
在 帮助中心 和 File Exchange 中查找有关 Deep Learning Toolbox 的更多信息
产品
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!
Translated by 
