[WEEK -5 ] Sound of The City

Authors: M.Çağdaş ÇAYLI , Buğrahan AKBULUT

This week we finally get some concrete results from our program,we have tried Neural Networks(NN) and Support Vector Machines(SVM) techniques on 8750 training data.In training;

We separate our data into two parts after shuffling our data matrix, first part for training (8000 *.wav sounds ) second for validation (750 *.wav sounds). Next we pass our training matrix to both SVM and NN.Because SVM algorithms are not scale invariant, we scaled each attribute on our data matrix to [-1,1] interval - We do this process for both train and validation data matrices to obtain meaningful results-. But we did not use this approach for NN because of its characteristics.

In validation part;

Our NN accuracy results with different activation functions,

Neural Network Results

Among these functions logistic function is given the best results, so there are some different results for logistic function below.(we played with number of neurons, number of hidden layers, learning rate and batch size)

number of neurons .1

number of neurons .2

number of hidden layers

batch size 1.1

batch size 1.2

learning rates

SVM results with different kernel functions and C values;

Note : In the table below all results are taken by using “ovr” as decision function! (We have also tried “ovo” but results did not change a lot.)

SVM Results

So we get the best results from NN we will use it later experiments and upgrades for our project.

Some Useful Resources

1.4. Support Vector Machines - scikit-learn 0.19.1 documentation

http://scikit-learn.org/stable/modules/neural_networks_supervised.html