Joseph Geumlek

Joseph Geumlek

Data Scientist at Kaiser Permanente. Interests are in machine learning, with a focus in privacy preserving techniques such as differential privacy. I am also interested in teaching and science communication. Outside of work, I enjoy video games and other such interactive fiction.

CSE 150A Intro to Artificial Intelligence: Probabilistic Reasoning (Remote)

undergraduate course, University of California San Diego, 2020

Remote instruction of an undergraduate course during the summer sessions of 2020.

Background

Following my successful instruction of CSE 150A in the winter of 2019, I again taught as the instructor of record in the summer of 2020. This places it after receiving my PhD, and was an appointment as a Summer Lecturer for UCSD.

The 10 week course was adapted into 5 weeks, and significant changes were made to accomodate remote synchronous and asynchronous instruction during the 2020 pandemic.

Curriculum

This course focuses on probabilistic modelling, using belief networks. A heavy emphasis is placed on following probility axioms to their logical conclusions in order to implement models to complete useful tasks. The start of the course also contains a review of probability basics.

No topics were removed from the 10 week version of the course, but assignments were shortened. The biggest omission was the removal of the last homework assignment, which is usually an open-ended mini-project of the student’s choosing.

  • Using Probability to Handle and Track Uncertainty.
    • Intepreting Probabilities as Beliefs.
    • Conditional Independence and Conditional Dependence (d-separation).
  • Inference on Models to Accomplish Tasks.
    • Defining Belief Networks.
    • Computing Conditional Probabilities on Belief Networks.
    • Enumeration Strategy for Inference.
    • Variable Elimination Strategy for Inference.
    • Naive Bayes Model.
    • Markov Chain Model.
    • Viterbi Algorithm for Hidden Markov Models.
  • Learning Parameters to a Model from Data.
    • Data Collection.
    • Maximum Likelihood Estimation.
    • Training Error.
    • Different Parameterizations of a Model.
  • Learning Parameters to a Model from Incomplete Data.
    • Hidden Variable Models.
    • Partially Observed Data.
    • Expectation Maximization.
    • Forward-Backward Algorithm for Hidden Markov Models.
  • Using Reasoning to Choose Actions and Strategy.
    • Markov Decision Processes.
    • Policy Iteration.
    • Value Iteration.
    • Q-Learning and Reinforcement Learning.
  • Bonus Topics (as Time Permits)

Presentation Style (and Reflecting Upon It)

This course was presented in a fashion to let students choose between synchronous and asynchronous instruction. Due to timezones, synchronous instruction was not feasible for a sizable number of students. And even for students in the Pacific time zone, the morning lecture times do not necessarily fit nicely into the lifestyles impacted by a global pandemic.

Following best practices for online course design, the content was placed on the Canvas LMS orgarnized into modules per week of the course. (More specifically, half weeks.) This clearly communicated the pace of the course. Within these modules, the lectures were split into a series of shorter, focused videos, averaging 6 minutes in length. Slides were also posted. Asynchronously watching the lectures, completing assignments, exams, and learning reflections was sufficient to pass the course.

For synchronous students, I hosted “live” lectures on Zoom, which consisted of broadcasting the videos, running in class polls and peer instruction activities, and responding to student questions. I was also able to digitally draw on and annotate the concepts in the videos in real time. These sessions were recorded and made available for students as well.

At the time of writing, it is still a bit early to tell how this method fared against other remote instruction efforts at UCSD. However, I am proud of my work, and students expressed that this style was quite effective.

It was designed to try to capture the power of digitally distributing the lecture in smaller chunks, while also provided students with a regular class time and routine for progressing through the material at a steady, instructor-moderated pace. There are a lot of extant methods for online instruction, but I still feel this approach offers a solid middle ground for instructors and students still adapting to online education.

Following from my experience in 2019, I still feel that flipped classroom settings, which also naturally translate to online education, do not adequately serve all learners equally. For students accustomed to in-class lecturing, I feel that empowering the usual familiar dynamics between learner and instructor is valuable. This year has been a stressful time. However, it also important to not treat these lectures like normal lectures: talking at students from the front of a room is already a questionable way of conveying knowledge, and its drawbacks are exacerbated by lengthy video conferencing.

The built in polling feature of Zoom was incredibly useful, providing almost minimal barriers to soliciting responses from engaged students without needing any additional devices. The break out rooms did serve as a replacement for discussing with nearby peers, but unfortunately it does not match the feedback provided by “reading the room” while lecturing in person.

Canvas also worked well for this course, but its module system did have drawbacks that I was unable to mitigate.