DLP_HW_1_Choijamts Bataa
.docx
keyboard_arrow_up
School
Syracuse University *
*We aren’t endorsed by this school
Course
691
Subject
Industrial Engineering
Date
Jan 9, 2024
Type
docx
Pages
3
Uploaded by PresidentLapwing3845
IST 691: Deep Learning in Practice
Homework 01
Instructions:
Answer the following questions in no more than one page per question. In addition
to the accuracy of your responses, the clarity, coherence, and concision of your writing are
critical factors to earning full credit for this assignment.
Cite
any source you use outside of
lecture notes. This includes the textbook. Reproducing or even paraphrasing an answer from a
generative AI tool, such as ChatGPT, is not allowed on this assignment. Submit your responses
in single Word or pdf document.
1.
In traditional machine learning (non-NN methods), we use feature engineering to
model complex relationships between observed variables (features) and a target
variable (response). When using deep learning methods, should we design and
incorporate feature engineering processes? Explain why or why not.
When using deep learning methods, there is no need to design and incorporate feature
engineering process because deep learning creates rules algorithmically while feature
engineering requires great amount of human intervention (Deep Learning in Visual
Approach, Chapter 1, p4).
2.
Explain in words or equations why we should introduce nonlinearity in neural
networks.
When fully connected network applies only addition and multiplication or linear
function, it collapses. To prevent this collapse, activation functions with nonlinear
operations are used (Deep Learning in Visual Approach, Chapter 13, p331).
3.
You are training a deep learning model to predict sentiment of Twitter posts — the
model predicts whether a post is “happy” or “sad”. Your model achieves 0.95
accuracy on the dataset you used to train the model. But when you take new posts
from Twitter and use your model to predict the sentiment, the model performs much
worse. What might have happened? What should you do to improve your model?
The model showed excellent performance on training dataset, but it underperformed on
new data, which means that there is overfitting. It might have happened that the model
used specific information in the training dataset rather than learning general rules. We can
improve the model by finding the moment when rules apply specific data and stop learning
at that moment. We can also improve the model by using regularization techniques. (Deep
Learning in Visual Approach, Chapter 9, p196).
4.
The MNIST dataset consists of images of dimension 28x28 pixels with one color
channel (28x28x1), with each image corresponding to a label between 1 and 10. To
build a classifier, we implement a multi-layer perceptron model with 3 hidden layers.
The first two hidden layers have 100 perceptrons each, and the third hidden layer has
30 perceptrons. Calculate how many weights will be updated for each iteration of
gradient descent. Show your work.
To calculate how many weights will be updated for each iteration of gradient descent in a
IST 691: Deep Learning in Practice
Homework 01
multi-layer perceptron (MLP) model, I will use the notion that each neuron is connected to
all neurons in the previous layer and each connection has its own weights (Deep Learning
in Visual Approach, Chapter 13, p328).
Firstly, in terms of input layer, the image has 28x28 neurons and there is no need for
weights. Next, there are three hidden layers, and first hidden layer has 100 perceptrons.
Each perceptron is connected to all 28x28 neurons in the input layer. In terms of the
second hidden layer, it also has 100 perceptrons, each connected to all 100 neurons of the
first hidden layers. For the last hidden layer, it has 30 perceptrons, each connected to all
100 neurons of the second hidden layer. Lastly, in terms of the output layer, there are 10
neurons, each connected to 30 neurons of the 3
rd
hidden layer.
Therefore, the total number of weights updated during each iteration of gradient descent is
100 perceptrons * 28x28 weights+100 perceptrons * 100 weights +30 perceptrons * 100
weights + 10 neurons * 30 weights = 797,300 weights
5.
Answer the following questions based on a close reading of this article and possibly
additional research (remember to cite your sources).
https://www.nytimes.com/2023/06/28/technology/facial-recognition-shoplifters-britain.html
a)
What is the technology being discussed in this article? How does it relate to
deep learning?
In this article, a solution on how security cameras are applied to combat petty crime
is discussed. When shoplifters enter a store, the face recognition software
recognizes, and alerts based on data from security camera. Way how the AI tool
works is simply about deep learning.
b)
What are some ethical concerns discussed in the article? Do you share these
concerns? Why or why not?
Facewatch included people’s biometric data into software without informing them or
making proper investigation on what actually had occurred. Due to this practice,
woman who is purchasing milk was recognized as barred shoplifter. Also, it
categorized problematic customers the same as shoplifters. Actually, they are
different.
c)
Are these concerns addressed by the company deploying this technology? How?
Facewatch amended its policy on when faces of offenders are added into the
program. It used to add offenders who even did single minor offenses, but the
company changed its policy and as a result, they would share information of only
serious and violent offenders and broadcast alert only repeat offenders.
Your preview ends here
Eager to read complete document? Join bartleby learn and gain access to the full version
- Access to all documents
- Unlimited textbook solutions
- 24/7 expert homework help