Bayes Rule and Probability Review (3)

Section Review

• Rules of probability, Bayes’ rule
• Distribution: conditional, joint, marginal
• Maximum likelihood estimation
• Note the distinction between probability and statistics
  • Probability : we’re given the parameteres, we want to study the distributions
  • Statistics : We’re given the data, we want to estimate the parameters (and also, test hypotheses)
• Point Estimates
  • when doing MLE, we end up with a single value
  • Suppose we have two estimates(e.g. CTR A vs CTR B)
    • How can we sure A is better than B?

Frequentist to Bayesian

The Bayesian asks, “instead of a point estimate, what is the distribution?”

• Frequentist : $\hat{\mu}$
• Bayesian : $P(\mu | data)$

From a distribution, we can find intervals (not called confidence intervals)

We can ask : “What is the probability that A is better than B?”

• Frequentist : point estimate, fixed but unknown parameters
• Bayesian : Those parameters are not fixed, but are also random variables
• and it’s why we use data to estimate the distribution of random variables

The Bayesian asks, “instead of a point estimate, what is the distribution?”

• Frequentist : $\hat{\mu}$
• Bayesian : $P(\mu | data)$

From a distribution, we can find intervals (not called confidence intervals)

We can ask : “What is the probability that A is better than B?”

• Frequentist : point estimate, fixed but unknown parameters
• Bayesian : Those parameters are not fixed, but are also random variables
• and it’s why we use data to estimate the distribution of random variables

 

Misconceptions about Statistics vs. ML

Definition of Machine Learning

• ML is when we build a model where the parameters of the model are learned from data
  • Linear Regression also uses MLE
  • Deep nueral networks also use MLE

• Statistics and ML are not as clear as you thought

• What is reward signal ? maximize clicks, donations, purchases, etc. (a.k.a. rewards)
• Recall the application (ex. Online advertising)

!! We want to build a model that maximizes rewards !!

 

 

Broad Definitions of Machine Learning

• ML is when we build a model where the parameters of the mdoel are learned from data
• Online learning
  • Static methods (e.g. neural networks) train once on a dataset, then never change
  • Online learning acts in real-time
    • Data is ingested one sample at a time, parameters are updated each time
  • The algorithm becomes smarter and smarter for each sample it collects
• There’s no need to name. Don’t try to distinguish statistics and ML.