Why is my CNN overfitting?

I'm trying to train a convolutional neural network (CNN) to classify raw ECG signals. Successes indicate that defibrillation worked on the patient and failures indicate that defibrillation did not work. I've included the .mat file with all the data. The matrix vf represents all the signals while the string array, R, represents the labels. vf has 2078 rows for each patient and 760 columns for samples per signal.

My training plot looks like a classic case of overfitting and I have included the image below. My validation accuracy is consistently lower than my training accuracy and the validation loss increases over time. I've tried to introduce batch normalization, L2 regularization, and data augmentation into my neural network to no avail.

Are there errors with my neural network layers? It seems that no matter what I do, the validation accuracy will not go higher than 55%.

Thanks in advance.

load cnnData

miniBatchSize = 128;
maxEpochs = 150;
learnRate = 0.001;

out = mat2gray(vf); % Min-max normalizes signals
data = reshape(out,[1 size(out,2) 1 size(out,1)]);
labels = categorical(R);

% Determining indices of training, validation, and test data
[trainInd,valInd,testInd] = dividerand(size(out,1),0.7,0.15,0.15); 

% Gathers training data and corresponding labels
XTrain = data(:,:,1,trainInd);
YTrain = labels(trainInd);

% Gathers testing data and corresponding labels
XTest = data(:,:,1,testInd);
YTest = labels(testInd);

% Gathers validation data and corresponding labels
XValidation = data(:,:,1,valInd);
YValidation = labels(valInd);

% Layers of Neural Network
layers = [imageInputLayer([1 size(out,2) 1])
    
    convolution2dLayer([1, 15],16,'Padding',1)    
    reluLayer 
    dropoutLayer(0.4)
    maxPooling2dLayer(2,'Stride',2)
    
    convolution2dLayer([1, 15],32,'Padding',1)    
    reluLayer 
    dropoutLayer(0.4)
    maxPooling2dLayer(2,'Stride',2) 
    
    convolution2dLayer([1, 15],64,'Padding',1)
    reluLayer
    dropoutLayer(0.4)
    maxPooling2dLayer(2,'Stride',2) 
    
    fullyConnectedLayer(2)
    softmaxLayer
    classificationLayer];

% Training Options of Neural Network
options = trainingOptions('adam',...
    'MiniBatchSize',miniBatchSize,...
    'MaxEpochs',maxEpochs,...
    'InitialLearnRate',learnRate,...
    'ValidationData',{XValidation,YValidation},...
    'Plots','training-progress',...
    'Shuffle','every-epoch',... 
    'Verbose',true); 

% Trains network and classifies test data
net = trainNetwork(XTrain,YTrain,layers,options);
Pred = classify(net,XTest);

% Creates confusion matrix
confMat = confusionmat(YTest,Pred);

[Matt J]

I've tried to introduce batch normalization, L2 regularization, and data augmentation into my neural network to no avail.

I think we need more details on the sweep of these parameters that you did. What range of regularization parameters did you test?

It is interesting that your validation loss climbs even though the accuracy stays about the same.

[Jared Stinson]

I tried including batch normalization layers before each relu layer. That didn't help so I tried moving the layers to after the relu layer and then after the dropout layers. None of those worked for me so I stopped trying to use batch normalization (also I read somewhere that batch normalization is not as effective when using dropout layers).

I included the L2 regularization option in the training options and tested the CNN with the L2 value from a logarithmic scale from 0 to 1 (0.1, 0.01, 0.001, 0.0001). None of those improved the validation accuracy though.

My data augmentation wasn't too complicated. I just used the flip function to flip each signal about the y axis and used those signals as extra data. Admittedly that helped a little bit as the validation accuracy kept up with the training accuracy up until the 10th epoch.

[ytzhak goussha]

You can try reducing the number of filters or their sizes