Suppose we model a signal with neural network or fuzzy models.
Does anyone have an idea for feature extraction from the constructed model??
for example we have N signals. If we construct AR or ARMA model for each of them, the coefficients of the model can be used as features for them. But in fuzzy or networks methods It does not seem to be useful and reasonable.
Thanks all
You mentioned coefficients of parametric models. However you can always use non-parametric signal features. E.g power spectra, cross-spectra coherence etc.. which should be applicable to any simulated signals. Actually maybe even estimators of these features calculated from parametric models(AR, ARMA) instead of just coefficients might be also useful.
Fuzzy methods are so diverse in practice, most of these are non-parametric . Feature Extraction is a sensitive task. One can use different mathematical/geometrical/signal processing routines to derive a meaningful feature of a signal. To me the question is so ambiguous.
Hi Mohammad. What exactly do you mean? You mean extracting the modeled features form a fitted neueral network (trained on an autoregressive time series)? If that's what you mean then you can fit the networkm and then do a sensitivity analysis by variying the input values and oserving the change outputs. This will tell you the relative sensitivity of the outputs to each input, their interactions, and hence allow you identify the relative importance and what features were relevant. Hope this helps, Sven
Hi Bagher, ;-)
Neural networks could be used for feature extraction. But for this purpose, instead of using their weights as feature, neurons of a middle layer (not output layer) is used as features. For more information you could google for "bottle neck neural networks".
In fuzzy field, I think you can use the parameters of membership function of tranined fuzzy system as features. for example you can use mean points of them. this approach is common when using GMM.
- Badges
- Science topic