2022-01-26

Lead Scribe: Lily

Admin

• sorry about notes

• private github repo –> Spring 2022

• grading contract FYI

  • Will be given further instructions on ways to achieve an ‘A’ or ‘B’

    • To get a ‘B’ you will only need to complete the paper and presentation

    • To get an ‘A’ you will implement a project (translation)

  • Paper and presentation will be assigned

    • Paper –> CS Conference Style

    • Draft due: last day before the presentation, will be posted for the class to review

Opening Question

What kind of data are you most in working with?

• Class response:

  • GIS data

  • Linguistic data (tweets, reddit posts)

  • Numerical data (tabular)

  • Video/Image

  • Time series

  • EHR/Medical related data

  • NLP

  • tabular/survey

How to Read a Paper

Model Based ML

• Discrete probabilities (distributions introduced in murder mystery chapter)

• Bernoulli

• Priors (probablistic guess about a random variable)

  • Are useful for working with less data to create strong inferences

    • Working with things when not a lot of data is available

  • Assumptions, expressed in a probability distribution

• Posterior

  • Inference given regularizer: Likelihood…

  • Most common posterior probability distribution we’re doing: Probability of parameters given data

• Point Estimate

  • This are the single values produced after training (weights)

  • Posterior mean

• Most of the probability distributions we’ll use belong to the exponential family

  • https://en.wikipedia.org/wiki/Exponential_family

• Conditional Probability

  • One for each value of the conditioning variable

  • (e.g.) Murder mystery –> murderer variable can be Grey or Auburn

• Marginal probability

  • (Section 1.2 – A Model of Murder)

  • “Probability of one event in the presence of all (or subset) outcomes of the other random variable…” (https://machinelearningmastery.com/joint-marginal-and-conditional-probability-for-machine-learning/)

• Maximum Likelihood Estimation

  • Assume a distribution, our goal will be to find the theta (parameter)

  • Maximizing, find parameters that will give us the highest probability (finding the one–parameter–that fits best)

• elicitation - an interdisciplinary field in statistics and psychology; study of how to get an expert’s distribution for how likely an event is to occur.

Prepare for next class

• Order of the weekly topics may change

• Dr. Brown will present next week, but we’ll start rotating the following week

• There are (2) readings, bring questions and prepare

Learning & Evaluation

• Read through the whole Learning and Evaluation Page after I post a notification to, there are some fixes to be made

• Bring Questions to class next week

• Be ready to work on your grading contract

Reading

The Scientific Method in the Science of Machine Learning and Value-laden Disciplinary Shifts in Machine Learning