TL;DR

• Extremely difficult
• 20-25 hours per week on average
• Enormous curve, 40% got me a B

Quick tips

• You don’t need the textbook
  • If you’re interested in the theory go for it, but it doesn’t help much for assignments
• Can work ahead in this class

What is Machine Learning?

Machine learning is getting computers to make decisions without being explicitly programmed to

Graded course material

• 2 Exams
  • Closed-notes
  • Each exam is worth 25% of your final grade
  • Difficulty level: Medium, they ask open-ended questions, and all answers can be found in the lectures (watch them for sure)
• 4 Projects
  • Difficulty level: Medium
  • Tedium level: Hard
    • I say this because they are very “run these 20 experiments”-esque, which just takes forever to code
  • Writeups: The writeups are very heavily scrutinized, and need to be really well written
    • 45/100 was my average score for writeups
    • Code is not king in this class, writeups are 👑

Project overview

Supervised Learning

This assignment in particular has you use any code package/library to “implement” the following:

• Decision trees with pruning
• Neural nets
• Boosting
• Support Vector Machines
• k-nearest neighbors

**PRO TIP** Use packages like pytorch, or Weka to do the heavy lifting for you
You do not need to code everything from scratch (it even says so in the project spec)

15% of your grade

• Find 2 interesting classification datasets:
  • Can be from anywhere, as long as they are non-trivial and not overly-complex
    • University of California-Irvine has some cool public datasets
    • Mine were an MLB baseball statistics dataset, to try and predict the MLB world series winner
    • And an obesity dataset, where you try to guess the obesity level of a person given life characteristics
  • At least 1 should be large
    • Mine had 112 and 2000+ records respectively
  • You will use these in project 3 for:
    • K-means
    • Independent Component Anlaysis (ICA)
    • Insignificant Component Analysis (PCA)

**PRO TIP** At least one dataset with more than 2 output classes is best for K-means, ICA and PCA

• Writeup tips:
  • Analyze the behavior of the models
    • Don’t just print the results, really go in-depth here
    • Why did the models behave the way they did?
  • Analyze your datasets in-depth
    • How often do output classes occur?
    • Is there any correlation between variables and output classes?
    • What’s different about each set what are the challenges with each set?
  • Metrics
    • Why did you use the error metric that you used?
    • Were there other error metrics you looked at?

**PRO TIP** Use a fancy paper template to write your paper, it will get you more points

Randomized Optimization

This project is very time consuming, because you run a total of 16 experiments

10% of your grade

Here you will experiment with:

• Randomized hill climbing
• Simmulated Annealing
• Genetic Algorithms
• MIMIC

**PRO TIP** Watch the lecture videos for this section to understand the algorithms
They are explained very well in the videos, and MIMIC is Dr. Isabell’s graduate thesis

• Find 3 problems
  • Bit-string representation problems are your friend, such as 4-peaks and k-colors

• Going back to project 1
  • You will use:
    • Randomized hill climbing
    • Simulated Annealing
    • Genetic algorithm
  • to train weights for the neural network from project 1 (Supervied Learning)
    • This part was honestly the most confusing, and I suggest doing it first, because it took the longest

• Writeup tips
  • Discuss hyper-parameter tuning
  • Discuss gradient descent performance

Unsupervised Learning

Insanely tedious, you do 20 experiments so ❗️start early❗️

• 2 clustering algorithms:
  • k-means clustering
  • Expectation Maximization

• 4 dimensionality algorithms
  • PCA
  • ICA
  • Randomized Projections
  • Any other feature selection algorithm you want

You’ll apply the 2 clustering algorithms to your 2 datasets, then, for each new clustering, you’ll apply all 4 dimensionality reduction algorithms.

So 2 x 2 x 4 = 16 😱

Then, you’ll run some stuff on your neural network from project 1 again

So 16 + 4 = 20 experiments total

Then you have to write about all of them 🥲

I have no advice, other than start early

I got 23/100 so good luck!!

Markov Decision Processes

This one is the most fun in my opinion

Markov Decision Processes is taking stat-action pairs, and training agents to do a task

Lunar Lander is one example of an MDP, but you won’t do that in this class
– CS 7642 Reinforcement Learning you do

• Find 2 MDP problems
  • Not overly complicated
  • One “small” and one “large”
    • I did Cart-pole and Taxi

• Value iteration vs policy iteration
  • To this day I don’t remember the difference
  • But you do both on the problems

• Reinforcement learning algorithm
  • Q-Learning is one popular choice

**PRO TIP** If you’re low on steam, I would invest more energy into this project, as it’s 15% of your grade and the most fun in my opinion

• Writeup tips:
  • Graph value and policy iteration convergence side-by-side somehow, either through time spent, iterations spent etc.
  • Do value and policy iteration come to same answer? Why or why not?
  • How do the sizes of the problem affect the algorithms?

Closing statement

While this class is tough, you can definitely get a B if you stick with it

Don’t worry if you’re getting 30’s/40’s/50’s on assignments, just do your best and it will be okay 🙂