Scribd
Upload
8 views

Homework 2

Uploaded by

mertdene10
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
8 views

Homework 2

Uploaded by

mertdene10
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

CMPE 442 Assignment #2

Due Friday, May 19, 23:59.


Note: The assignment is to be done individually. You are going to submit the report and
code separately. Please do not include codes in the report. Your report should be self-
contained, i.e. explain what you are doing, put the resulting graphs, the
hyperparameters that you use, and write your observations.

Note: This assignment worth 10% of overall course grade.

In this assignment you are going to implement a neural network with one hidden layer. First, you
will implement a NeuralNetwork class and then use that network to learn digits dataset.

1- [60 pts] Implement a NeuralNetwork class according to the template that is provided to you
(NN_template.ipynb). In this template you will find a class NeuralNetwork with a constructor
and 4 methods. Some of the methods are partially filled. Feel free to add you own code where
you think appropriate.

a) [30pts] Fill the feedforward() method

b) [30pts] Fill the backprop() method. Ignore the regularization part in the algorithm given
in slides (i.e. λ=0).

2- [40 pts] Train your network on digits dataset. It is a toy dataset in sklearn. The template
shows the way of fetching the data and splitting it into three sets: train, validation, and test
sets. Note: Do not touch test set until the last step of your assignment.

Start with default number of units in the hidden layer as 64 (sl2=64). Fix the number of
epochs to 100, i.e. iterNo=100, throughout your assignment.

a. [15 pts] Try different learning rates = {0.01, 0.1, 0.5, 0.9}, This means that you train four
networks and for each network that you learn test it on validation set. Report on the
learning rate that give you the best performance on validation set.

b. [15pts] Try different values for 𝑠𝑙2 = {16, 32, 64, 128}. The learning rate is fixed to the
one that gave you the best result in (a). As in the previous case, train four networks and
for each network that you learn test it on validation set. Report on the number of hidden
units that give you the best performance on validation set.

c. [10pts] Re-train the network combining train set and validation set, using the best
hyperparameters that you found in (a) and (b). Finally, test the model on the test set.
Report on the accuracy and confusion matrix on test set.

You might also like