Skip to main content

Machine Learning by Shalmali Joshi

Framing​

My Causal Inference class had a Turing quote: “What we want is a machine that can learn from experience.” One of Dr. Joshi’s slides mentioned a Herbert Simon quote about learning itself — how you improve performance from experience. Gave me something to think about.

What ML Is​

ML is how you improve performance (P) at some task (T) with experience (E) — a <P,T,E><P, T, E> tuple of sorts.

Typically, when we use computers to perform tasks:

Inputs/Data←operate uponPrograms→produceOutput\text{Inputs/Data} \xleftarrow{\text{operate upon}} \text{Programs} \xrightarrow{\text{produce}} \text{Output}

This is still the case in ML, but the outputs are not data in the usual sense. Rather, the output is a model which you can use to operate upon other/new input data to yield insights, make predictions, and generate knowledge.

ML is good for when human expertise doesn’t yet exist.

Flavors of Learning​

Learning itself in ML comes in various flavors:

  • Supervised — involves experts
  • Unsupervised
  • Semi-supervised — a blend of the two
  • Reinforcement — involves experts with cookies for the machine if it’s doing a good job

Under the covers, it’s all statistics, and so these learning flavors can be applied in several contexts:

  • Descriptive — what is
  • Prescriptive — what should be
  • Predictive — what will be
  • Generative — new stuff

Lecture Arc​

The lecture introduced these ML flavors in some depth and ended with a discussion on the Hot New Thing: GPTs.