INT305 – Coursework 1
Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit
INT305 – Coursework 1
|
Assessment Number |
1 |
|
Contribution to Overall Marks |
15% |
|
Submission Deadline |
16/10/2025 |
|
Coursework Submission Guideline Write the answers under each sub-task. That is to say, you need to copy the question first in word/latex, then write your answers below the copied question. After you answer all the questions, please transfer your word/latex into pdf and submit the final pdf. Submit file format: Single column; Font size: #12; Page number: no more than 15; No need to prepare a coversheet, you can directly copy the questions and write the answers. Submit file name: INT305-CW1-Name-studentID.pdf (e.g., INT305-CW1-SanZhang-2025305.pdf) |
Assessment Objective
This assessment aims at evaluating students at Having a solid understanding of the theoretical issues related to problems that machine learning algorithms try to address. And check if the students are able to ascertain the properties of existing ML algorithms and new ones.
Tasks
For a training example (xi, yi ) with K classes:
s = wxi = score vector sj (sj = predicted score for class j)
Task 1: SVM Loss (35’)
1) Derive SVM loss
for xi ; set Δ as the margin hyperparameter. (1’)
2) Derive the gradient
for
k = yi (correct class) (5’)
k ≠ yi (incorrect class) (5’) And explain:
why is Δ only applied to incorrect classes? (3’)
How does Δ enforce a “safety margin”? (5’)
3) Given scores S = [3, 一1, 4] for classes [“cat”, “dog”, “bird”]; True class yi is “cat” (index=0); Δ= 1, compute and explain:
for each score (3’)
How does
change if Δ= 2 (5’)
4) Using the same score S = [3, 一1, 4] and yi = 0:
Compute
(3’)
Interpret the gradient: Why are some values positive, negative, or zero? (5’)
Please show the step-by-step process in the report for all the subtasks.
Task 2: Softmax Loss (35’)
1) Derive
softmax probability pj for class j (1’)
softmax loss
(1’)
2) Derive the gradient
for
k = yi (correct class) (5’)
k ≠ yi (incorrect class) (5’) And explain:
Why does minimizing
force pyi → 1? (3’)
3) Given scores s = [3, 一1, 4] for classes [“cat”, “dog”, “bird”]; True class yi is “cat” (index=0), compute and explain:
pj (3’)
for the given scores (3’)
What happens to
if all scores are scaled by 2 (i.e., snew = 2s) (4’)
4) Using the same score s = [3, 一1, 4] and yi = 0:
(3’)
How does the gradient for the correct class differ from SVM? (7’)
Please show the step-by-step process in the report for all the subtasks.
Task 3: Comparative Analysis (30’)
1) Using the same score s = [3, 一1, 4] and yi = 0:
Compare the previous derived losses
and
, which is larger? (2’)
Why? (3’)
2) Using the gradients from task 1 and task 2:
How does SVM penalize “near-miss” errors (e.g., sj ≈ syi) vs. “large-margin” errors? (2’)
How does Softmax adjust probabilities for low-confidence predictions? (3’)
3) Why is SVM loss called “max-margin” and Softmax “cross-entropy”? Please use your own words to define them and show your understanding of the definition and meaning of them. (8’)
4) Compare between SVM loss and Softmax loss, which is better and why? (in this case, please not just focus on previous tasks and show your understanding of the two losses) (12’)
Please show your analysis in the report for each subtask.
2025-10-13
Machine Learning