quartz/content/notes/17-ML-in-IA-1.md

title	aliases	tags
17-ML-in-IA-1		comp210 lecture

traditional approaches

• hand crafted and curated
• based on intuition not evidence (e.g., password criteria example)
• static and difficult to scale and adapt

Everything we do generates data. This can be used to learn what works and what doesn't

We use the data to evolve from observations to models

data doesn't lie. humans do (sometimes using data)

ai uses machine learning, data science applies machine learning. statistical learning is machine learning

ML

using mathematics and algorithms to build models from data for prediction and insight

key concepts

• getting dta
• feature enginering
• model testing/validation
• knowledge extraction

supervised vs unsupervised learning — often feed into each other

feature engineering

converting domain variables into actionable features

e.g.,

• edges from images
• word frequencies in documents
• converting TCP stream into counts of particular packet types

insights

modelling uncovers relatonships within inderlying process represented by the data

e.g., detecting a password attack

benefits

• provide insight
• automattion
• adapt to changes
• scale larger and more complex problems

threats

need to secure all of them - smartly