Signal processing for machine learning toolbox

Has anyone used so far toolbox

https://www.mathworks.com/videos/signal-processing-for-machine-learning-99887.html ?

I have an assignement to classify time series signals from my own data set.

I collected my own measurements with Myo armband and organized them according DataPreparation.m and PrepareData.m functions. I have 5 different classes of arm movements (drinking water, eating, moving glass, opening drawer and scratching). When I was labeling the data I used column vectors with a) 32 samples, b) 64 samples. I had 20 subjects and more than 50 labels per different kind of movement/ per subject.

I used the highpass filter from the toolbox and I calculated features with extractAllFeatures.m. I used ffd neural network with 18 neurons in hidden layer.

In both cases I got confusion matrix with clasiification accuracy of slightly over 50%.

Why I don't get better clasiification accuracy when the dataset is twice bigger (column vector of 64 samples compared with column vector of 32 samples)?

Should I make changes in calculation of features and in which direction?

I can attach all necessary files- confusion matrices, signal in time domain, changes of the functions of original toolbox etc.