AAI/CPE/EE 695 Applied Machine Learning: Midterm Exam
Spring 2024
This midterm contains two questions, each of which requires you to solve a coding problem,
then discuss the results. You are free to use library implementations of methods and tools
from Scikit-Learn or PyTorch.
Make sure to answer every part of each question.
For this exam, submit a single Jupyter Notebook (*.ipynb) file via Canvas. In addition to your
code, include answers to any discussion questions in markdown cells in the same file.
Question 1 (60 points):
Dataset: MidtermQ1Data.zip contains images of two types of fruit: green apples and oranges.
The images are divided into two sets: a train set, and a test set.
Setup: When applying machine learning methods to complex data, one of the most common first
steps is to extract salient features. For images, one such feature could be a color histogram,
which measures the occurrence rates of each color within an image:
from PIL import Image
import numpy as np
def get_image_histogram (image_filename, nbins=5):
im =
Image.open(image_filename)
im = im.resize((100,100))
im_array = np.array(im).transpose(2,0,1)
im_array =
im_array.reshape(3,-1)
im_array = np.transpose(im_array, (1,0))
return np.histogramdd(im_array, bins=nbins,range=[(0,255) for
in range(3)])[0].flatten()
Task: Use this color histogram feature to preprocess the images, then choose two of the
following methods and apply them to classify the images as either a green apple or an orange:
Method
Logistic Regression
Decision Tree
Support Vector Machine (SVM)
Random Forest
Hyperparameters
Maximum Iterations
Maximum Depth
Maximum Iterations
Maximum Leaf Nodes
Artificial Neural Network (a.k.a Multilayer Perceptron) | Learning Rate
1. For each of your chosen methods:
a.
Use Scikit-Learn's implementation to train the model on the train set, and employ a
grid search using cross-validation to find an optimized value of the listed
hyperparameter.
b. Use the best hyperparameter value you found and train each model on the full train
set 2. Compare the performance of your two chosen methods:
a. Plot the results of the grid searches: the hyperparameter value on the x axis, and the
performance metric on the y axis. Discuss the results of the grid searches and
possible reasons for the behavior exhibited by the systems.
b. Measure each method's performance on the test set; report Accuracy and F1 Score
on the test set.
C.
Discuss any difference in performance between the two methods and possible
reasons for their relative performance.
3. Explain whether the color histogram is a good or bad choice of feature for this task and why.
Question 2 (60 points):
Dataset: Gradesheet.csv contains my (anonymized) grading rubric sheet for Homework 1 of this
course. Student names are replaced by an "ID" column containing integers. Columns [1a...1f]
and [2a...2d] contain scores for each sub-question, and the column “points" contains the total
grade.
For each sub-question, I assign each student's response a score from 0 to 1 (inclusive). Each
sub-question is worth a certain amount of points. To get the student's total grade, I multiply each
of the scores by the corresponding amount of points, then add up these products.
The maximum grade was 60 points.
Each subpart of one question is worth the same amount of points; so {1a, 1b, 1c, 1d, 1e, 1f} are
each worth the same, and {2a, 2b, 2c,2d} are each worth the same.
point_value₁α = point_value₁b
point_value2a = point_value2b
=
=
= point_value1f
= point_value2c = point_value2d
Task: Using any of the learning methods we've covered in this course, figure out how many
points each sub-question is worth. Brute-force solutions or guessing will not be accepted as valid
solutions.
Discuss how you arrived at your solution, and why the method you chose makes sense for this
problem.
(Hint: Write out the equation to calculate a student's final grade; it's format should look familiar)
