Machine Learning Homework 9-Explainable AI Solution

Machine Learning HW9
Explainable AI


Outline
● Topic I: CNN (HW3)
○ Model & Dataset
○ Task
○ Lime
○ Saliency Map
○ Smooth Grad
○ Filter Visualization ○ Integrated Gradient ● Topic II: BERT (HW7)
○ Task
○ Attention Visualization
○ Embedding Visualization
○ Embedding Analysis
Topic I: CNN explanation
Model: Food Classification
● We use a trained classifier model to do some explanations
● Model Structure:CNN model
● Dataset: 11 categories of food (same dataset in HW3)
○ Bread, Diary product, Dessert, Egg, Fried food, Meat, Noodles/Pasta, Rice, Seafood, Soup, and Vegetables/Fruit

Task
● Run the sample code and finish 20 questions (all multiple choice form)
● We’ll cover 5 explanation approaches
○ Lime package
○ Saliency map
○ Smooth Grad
○ Filter Visualization ○ Integrated Gradients
● You need to:
○ Know the basic idea of each method
○ Run the code and observe the results
Task: Observation
● In this homework, you only need to observe these 10 images.
● Please make sure you got these 10 images in your code.
● In the questions, the images are marked from 0 to 9.
● We encourage you to observe other images!
Lime
Question 1 to 4
● Install the Lime package -> pip install lime==0.1.1.37
GitHub repo: https://github.com/marcotcr/lime
Ref: https://reurl.cc/5G8EGG
Saliency Map
Question 5 to 9
● Compute the gradient of output category with respect to input image.
Ref: https://reurl.cc/6ELeLk
Smooth Grad
Question 10 to 13
● Randomly add noise to the input image, and get the heatmap. Just like what we did in the saliency method.
Ref: https://arxiv.org/pdf/1706.03825.pdf
Filter Visualization
Question 14 to 17
● Use Gradient Ascent method to find the image that activates the selected filter the most and plot them (start from white noise).
Ref: https://reurl.cc/mGZNbA
Integrated Gradients
Question 18 to 20
● Flexible baseline
Ref: https://arxiv.org/pdf/1703.01365.pdf

Topic II: BERT explanation
Task
● Run the sample code and finish 10 questions (all multiple choice form)
● We’ll cover 3 explanation approaches
○ Attention Visualization
○ Embedding Visualization
○ Embedding analysis ● You need to:
○ Know the basic idea of each method
○ Run the code and observe the results
Attention Visualization
Question 21 to 24
● Visualize attention mechanism of bert using https://exbert.net/exBERT.html
Alternative link: https://huggingface.co/exbert/ Ref: https://arxiv.org/pdf/1910.05276.pdf
Tutorial: https://youtu.be/e31oyfo_thY
Embedding Visualization
Question 25 to 27
● Visualize embedding across layers of BERT using PCA (Principal Component Analysis)
● Fine-tuned for Question Answering
Embedding Analysis
Question 28 to 30
● Compare output embedding of BERT using:
○ Euclidean distance
○ Cosine similarity
You only need to change code in the section “TODO” !
Grading
● 30 multiple choice questions
● CNN: 20 questions
○ 0.3 pt for each question
● BERT: 10 questions
○ 0.4 pt for each question
● You have to choose ALL the correct answers for each question ● No leaderboards & reports are needed!!
Submission
● You can answer the questions unlimited times
● The length of anwsering time of the assignment is unlimited ● We will consider the latest submission as the final score
● Remember to save the answer when answering the questions!
Links
● Code: [Colab]
● Questions: [gradescope]
Please don’t change the original code, unless the question request you to do so.
If any questions, you can ask us via…
● NTU COOL (recommended)
○ https://cool.ntu.edu.tw/courses/11666
● Email