In this assignment, you will solve practical and interesting problems. By completing the project, you will gain valuable hands-on experience in the design, implementation and evaluation of classification algorithms. The details are listed as below.

Assignment 2
CPS584 – Advanced Intelligent Systems and Deep Learning
Released Date: 10/05/2022
Requirements
In this assignment, you will solve practical and interesting problems. By completing the
project, you will gain valuable hands-on experience in the design, implementation and
evaluation of classification algorithms. The details are listed as below.
You are provided with the “DogCat.zip” file which contains images of two classes: Dog and
Cat. For each class, 25 training images and 15 testing images are given.
1. Extract the following features separately for each training and testing image:
a. Haar-like features (30-dimension as in Assignment 1). You should resize the
image to the size of 128 x 128 (not 48 x 28 as in Assignment 1) before extracting
Haar-like features.
b. Deep Learned Features at fc7 layer of AlexNet (4096-dimension). You should
resize the image to the size of 227×227 before extracting Deep Learned
Features. You may refer to Lab5 for AlexNet.
2. Use additional 30 training images (as you already collected in Assignment 1) for both
classes: Dog and Cat. Perform the K-nearest neighbor (KNN with K ϵ {1, 3, 5, 7, 9})
and Neural Networks (NN) on the testing images. Note that both classifiers (KNN and
NN) are trained on the new training data – there are 25 (already provided) + 30 (newly
collected) = 55 training images for each class. And please report the accuracy rate for
each class (KNN-Haar, KNN-Deep Learned Features, NN-Haar, NN-Deep Learned
Features).
3. Flip all training images in (3) to have 45 more training images for each class (as shown
in the example below). Please perform the K-nearest neighbor with different Ks (K ϵ
{1, 3, 5, 7, 9}) with the new training data. And please report the accuracy rate for each
class (KNN-Haar, KNN-Deep Learned Features).
FLIP
2
4. Use the same training set in (3). Concatenate Haar-like features and Deep Learned
Features for each training and testing image (30 + 4096 = 4126-dimension). Perform
the K-nearest neighbor with different Ks (K ϵ {1, 3, 5, 7, 9}). And please report the
accuracy rate for each class (KNN-Haar+Deep Learned Features).
5. Discuss the accuracy rates in (2), (3), and (4). For example:
a. Will more training data lead to a better performance?
b. Will different Ks have different results?
c. Will feature concatenation improve the performance?
d. Your own observations.
What to Submit
1. A well-documented MATLAB program that implements the aforementioned problem in the
Assignment 2. You must submit your program source code and the newly collected training
data set.
2. A well-written, concise project report. It should include: (a) title and names of group
members; (b) the analysis of each problem; (c) the issues during the implementation; (d) the
solutions to overcome the issues in (c); (e) the contribution of each individual member; and (f)
the powerpoint slides (maximum 20 slides) used in the Assignment presentation.
For each group, you must submit the files above in a single zipped folder. Your group will be
required to do a face-to-face evaluation for the grading.
Important: Your submission will be thoroughly checked. If any plagiarism (from Internet,
former students, or anywhere else) is found in this assignment, an F will be assigned to course
grade and an academic dishonesty report will be given.