convolution kernel size about master-thesis HOT 1 CLOSED

Hi there, thanks for the good question.

The neural networks you see are the result of experiments I did tuning the kernel sizes and number of filters employed in each convolutional layer.
I received a few hints to improve performances by tuning those values, you can see all in this DataScience StackExchange Q/A.

The input size is 2x128, so I thought that a 2x2 filter (that fits the size of the 2x128 matrix) would perform better than other ones. The value for the kernel size of the second convolutional layer (4x4) just came from tuning the values, train the neural network, interpretating the results and then choosing next values to test. I don't have an answer why even kernel sized filters perform better in this case, but results tells that they actually do.

from master-thesis.