How do I relate the scores fscmrmr to the feature names used in the network?

I am using an audio classifier with the following features

aFE = audioFeatureExtractor("SampleRate",fs, ...

"SpectralDescriptorInput","melSpectrum", ...

"spectralCentroid",true, "spectralSlope",true,"spectralFlux",true,...

"spectralKurtosis",true,"spectralSkewness",true,"pitch",true,...

"harmonicRatio",true,"zerocrossrate",true,"shortTimeEnergy",true,...

"spectralSpread",true,"spectralRolloffPoint",true,...

"spectralCrest",true,"spectralDecrease",true,...

"spectralEntropy",true,"spectralFlatness",true,...

"mfccDeltaDelta",true,"gtcc",true,"gtccDelta",true,...

"gtccDeltaDelta",true,"mfcc",true,"mfccDelta",true);

I use these features to train the following network

% Define the Deep Neural Network

layers = [ ...

sequenceInputLayer(numFeatures)

lstmLayer(100,"OutputMode","last")

fullyConnectedLayer(numel(unique(labelsTrain)))

softmaxLayer

classificationLayer];

% Define the training options

options = trainingOptions("adam", ...

"Shuffle","every-epoch", ...

"ValidationData",{featuresValidation,labelsValidation}, ...

"Plots","training-progress", ...

'InitialLearnRate',1e-3,...

"MaxEpochs",100,...

"Verbose",true);

% Train the Network

net = trainNetwork(featuresTrain,labelsTrain,layers,options);

Here is my question... How do I use the ouput of [idx,scores] = fscmrmr(XX,YY); to assess which features are the most important? I enabled 21 features in 'aFE' but the FeatureVectorLength is 93. How do I correlate the scores and indices returned with the fscmrmr(X,Y) funnction with features I defined in aFE?

Answer 1

[Brian Hemmat]

Many audio "features" are vectors. For example, the default MFCC feature has 13 elements. So if you just turned on MFCC, then you'd have a feature vector length of 13. You can get the mapping of audioFeatureExtractor output using the info method.

I think this example would answer your questions:

https://www.mathworks.com/help/audio/ug/audio-feature-selection-for-machine-learning-tasks.html

[Nicholas Pulsone]

Thank you! The info(aFE) command provides me with the information that I need,