Questions $(100$ points$)$: In this assignment, you will train feedforward neural networks with three

layers to classify motion data.

You can download "Human Activity Recognition Using Smartphones" data set from

https:$//$archive$.$ics.uci.edu$/$dataset$/240/$human$+$activity$+$recognition$+$using$+$smartphones link.

Then, unzip UCI HAR Dataset.zip in human$+$activity$+$recognition$+$using$+$smartphones$/$ directory

into the same folder.

In total, there are $10299$ movement patterns with $561$ features belonging to $6$ classes in the data

set. These features are explained in detail at human$+$activity$+$recognition$+$using$+$smartphones$/$

UCI HAR Dataset$/$features$.$txt file. This data set is divided into two as training and test subsets.

Training and test sets include $7352$ and $2947$ patterns, respectively.

You can benefit from "Human activity recognition with different artificial neural network based

classifiers" named conference paper in the

$"$

https:$//$iecexplore$.$ieee.org$/$abstract$/$document$/7960559"$ link to understand data set and

assignment better.

As a first step, load training and test data sets with similar codes to the following.

x$\_$train $=$ load $("$human$+$activity$+$recognition$+$using$+$smartphones$/$UCI HAR

Dataset$/$train$/$X$\_$train.txt$")$;

Y$\_$train $=$ load $("$human$+$activity$+$recognition$+$using$+$smartphones$/$UCI HAR

Dataset$/$train$/$Y$\_$train.txt$")$;

X$\_$test $=$ load$("$human$+$activity$+$recognition$+$using$+$smartphones$/$UCI HAR Dataset$/$test$/$X$\_$test.txt$")$;

Y$\_$test $=$ load $("$human$+$activity$+$recognition$+$using$+$smartphones$/$UCI HAR Dataset$/$test$/$Y$\_$test.txt$")$;

Here, Y train and Y test variables include class labels from $1$ to $6.$ You must transform

them into the one hot encoding scheme.

After that, you must generate a three$-$layer neural network, which takes $561$ features to the

first hidden layer consisting of $200$ neurons with hyperbolic tangent activation functions.

The second hidden layer of the neural network consists of $100$ neurons with rectified linear

unit activation functions. The output layer consists of $6$ neurons with either sigmoid or

softmax activation functions. Finally, cost function will be chosen either MSE or Cross

Entropy. To do these, you benefit from following codes.

h$1\_$size $-260$;

h$2\_$size $=160$; epochlimit $=80$;

input$\_$size train, $2$