CogSci131 Assignment 1 NeuralNetsSpring24

Assignment#1, Neural Networks

February 23, 2024

1 Khan’s Building in Bangla Desh and Starry Night

Figure 1.1: Figure Produced with Neural Net

1.1 Instructions.

Assignment #1 CogSci 131 Spring 2024

Neural Nets

Group Assignment

Total number of points 100.

Instructions

Please submit a working jupyter notebook with the solution to the questions below. Do

not use someone else’s work; all code has to be your own. Use the .ipynb scripts provided

in the modules section. You can submit a pdf �le in addition to your ipynb �le, or use the

markdown language in the script.

The deadline for this assignment is Wed. March 6 at 11:59 pm . You grade will be low if

you do not submit a working ipynb �le! So… watch out!

1.2 First Problem. Understand the structure of a neural network.50

Using the example found in neural nets demysti�ed and in the script provided, write a pro-

gram that minimizes the cost function to a given accuracy set in advance by you. Please

do not get stuck in the meaning of accuracy. This simply means a given threshold that

is reasonable and you think your program could reach, for example 0.1% of the target.

Notice that the program does part of the job for you. In this case there are two directions:

NN.Ŵ1 = NN.Ŵ1 + scalar

and the other direction:

NN.Ŵ1 = NN.Ŵ1 − scalar

and the same for NN.Ŵ2. J is the cost function.

Figure 1.2: Figure Produced with Neural Net

Plot the cost vs iteration and see if you get a similar �gure to the one I am showing above.

Scalar is a parameter that allows to change the step of the iteration. Some people call this

the learning rate for some mysterious reason. Explore what happens if the learning rate is

unusually large. I provide an example in the code.

1.3 Second Problem. Use ReLu instead of sigmoid activation. 50 Points

Use the ReLu activation function instead of the sigmoid function to construct a neural net

of the same dimensions used in 1.2. Compare how fast this neural network works with the

one that uses the sigmoid function activation. Add a new hidden layer and compare with

the previous network. Increase the number of hidden layer units to 10. Compare again.