Course learning outcomes#

In this course, we will learn some advanced machine learning methods in the context of natural language processing (NLP) applications. Markov chains, hidden Markov models, topic modeling, recurrent neural networks.

  • Carry out basic text preprocessing.

  • Specify Markov chains and carry out generation and inference with them.

  • Explain the general idea of stationary distribution in Markov chains.

  • Explain what are hidden Markov models and carry out inference and decoding with them.

  • Broadly explain Latent Dirichlet Allocation (LDA) approach to topic modeling and carry out topic modeling on text data.

  • Broadly explain Recurrent Neural Networks (RNNs) and different types of RNNs.

  • Broadly explain the idea of self-attention and transformers

Deliverables#

Click to expand!

The following deliverables will determine your course grade:

Assessment

Weight

Where to submit

Lab Assignment 1

15%

Gradescope

Lab Assignment 2

15%

Gradescope

Lab Assignment 3

15%

Gradescope

Lab Assignment 4

15%

Gradescope

Quiz 1

20%

Canvas

Quiz 2

20%

Canvas

See Calendar for the due dates.

Teaching team#

Click to expand!

Role

Name

Lecture instructor

Varada Kolhatkar

Lab instructor

Varada Kolhatkar and Vincent Liu

Teaching assistant

Daniel Ramandi

Teaching assistant

Harsh Sharma

Teaching assistant

Mohit Pandey

Teaching assistant

Ngoc Bui

Lecture schedule#

This course will be run in person. We will meet three times every week: twice for lectures and once for the lab. You can refer to the Calendar for lecture and lab times and locations. Lectures of this course will be a combination traditional live lecturing, class activities, and pre-recorded videos. Drafts of the lecture notes for each week will be made available earlier in the week.

This course occurs during Block 6 in the 2023/24 school year.

Lecture

Topic

Assigned videos/Readings

Resources and optional readings

0

Course Information

đź“ą

  • Videos: 16.1

    		•

    1

    Markov Models

  • Markov chains in action

    • 2

    Language models, PageRank, text preprocessing

    đź“ą

  • Videos: 16.2

    		•

  • OpenAI GPT3 demo
  • Dan Jurafsky’s videos on PageRank
  • Dan Jurafsky’s video on tokenization

    		•

    3

    Hidden Markov models

  • Nando de Freitas’ lecture on HMMs
  • A gentle intro to HMMs by Eric Fosler-Lussier

    		•

    4

    HMMs decoding and inference)

    (optional) HMM Baum-Welch (unlisted)

  • Nando de Freitas’ lecture on HMMs
  • A gentle intro to HMMs by Eric Fosler-Lussier

    		•

    5

    Topic modeling

  • Dave Blei video lecture, paper

    		•

    6

    Introduction to Recurrent Neural Networks (RNNs)

  • The Unreasonable Effectiveness of Recurrent Neural Networks
  • Highly recommended: Sequence Processing with Recurrent Networks

    		•

    7

    Introduction to Transformers

    8

    Applications of Transformers

    The labs are going to be in person. We will also be holding a short 1-hour parallel Zoom session for each lab run by the TAs so that people who cannot join in person have an opportunity to ask questions and get help. You will be able to access appropriate Zoom links via Canvas. There will be a lot of opportunity for discussion and getting help during lab sessions. Please make good use of this time.

    Installation#

    We are providing you with a conda environment file which is available here. You can download this file and create a conda environment for the course and activate it as follows.

    conda env create -f env-dsci-575.yml
conda activate 575

    In order to use this environment in Jupyter, you will have to install nb_conda_kernels in the environment where you have installed Jupyter (typically the base environment). You will then be able to select this new environment in Jupyter. If you’re unable to see the environment in Jupyter, you might have to install the kernel manually. See the documentation here. For more details on this, refer to your 521 lecture 7.

    I’ve only attempted to install this environment file on a few machines, and you may encounter issues with certain packages from the yaml file when executing the commands above. This is not uncommon and may suggest that the specified package version is not yet available for your operating system via conda. When this occurs, you have a couple of options:

    1. Modify the local version of the yaml file to remove the line containing that package.

    2. Create the environment without that package.

    3. Activate the environment and install the package manually either with conda install or pip install in the environment.

    Note that this is not a complete list of the packages we’ll be using in the course and there might be a few packages you will be installing using conda install later in the course. But this is a good enough list to get you started.

    Course communication#

    Click to expand!

    We all are here to help you learn and succeed in the course and the program. Here is how we’ll be communicating with each other during the course.

    Clarifications on the lecture notes or lab questions#

    If there is any clarification on the lecture material or lab questions, I’ll post a message on the course Slack channel and tag you. It is your responsibility to read the messages whenever you are tagged. (I know that there are too many things for you to keep track of. You do not have to read all the messages but please make sure to carefully read the messages whenever you are tagged.)

    Questions on lecture material or labs#

    If you have questions about the lecture material or lab questions please post them on the course Slack channel rather than direct messaging me or the TAs. Here are the advantages of doing so:

    • You’ll get a quicker response.

    • Your classmates will benefit from the discussion.

    When you ask your question on the course channel, please avoid tagging the instructor unless it’s specific for the instructor (e.g., if you notice some mistake in the lecture notes). If you tag a specific person, other teaching team members or your colleagues are discouraged to respond. This will decrease the response rate on the channel.

    Please use some consistent convention when you ask questions on Slack to facilitate easy search for others or future you. For example, if you want to ask a question on Exercise 3.2 from Lab 1, start your post with the label lab1-ex2.3. Or if you have a question on lecture 2 material, start your post with the label lecture2. Once the question is answered/solved, you can add “(solved)” tag before the label (e.g., (solved) lab1-ex2.3). Do not delete your post even if you figure out the answer on your own. The question and the discussion can still be beneficial to others.

    Reference Material#

    Click to expand!

    Online resources#

    Books#

    Policies#

    Please see the general MDS policies.