Page cover image

#1: Components of a Learning System

  • 3 min

#1: Components of a Learning System content

Created
Topics

Components of a Learning System

notion image

Example: Credit Approval System

Input: Customer Information (demographics, income, age, gender)
Output: Binary credit decision (made money or not)
Unknow target function f:x→yf:x→y (Idea credit approval formula)
↓\downarrow
Training Data D=[(x1,y1),...,(xN,yN)]D=[(x_1,y_1),... , (x_N,y_N)] (histoical records of current customers)
  • typically I.I.D (independent & identically distributed)
↓\downarrow
Learning Algorithm → Final Hypothesis g≈fg\approx f (Learned credit approval func)
↓\downarrow
Hypothesis Set (Condidate predictive functions)

A Simple Learning Model

Hypothesis Set: All linear functions HH

Weight w(t)

Update: w(t+1)←w(t)+y(t)x(t)w(t+1)←w(t)+y(t)x(t)
Assuming data linearly seperable: this algoritm h(x) works

Errors

Eout=E_{out}= out-of-sample error (unknown)
Ein=E_{in}= in-sample error (training error - known)

Hypothesis Space H\mathscr{H}

The set of functions that includes g that best approximates f
Picking a hypothesis space = selecting the type of algo / modal to use

Complexity of H\mathscr{H}

  • Complex H\mathscr{H} : better chance of approximating ff in-sample
  • Simple H\mathscr{H}: better chance of generalizing out-of-sample

Goal of Machine Learning

Minimize EoutE_{out} given DD
 

Learning Paradigms

Supervised Learning

Explicitly provided w/ labeled inputs

Online Learning

Does not have access to all data upfront
Given the algo one example at a time
e.g. auto driving

Active Learning

Algo allowed to query egs for its training

Unsupervised Learning

Only given input examples w/o annotations
  • Like human's obsersational learning
  • A hard problem

Applications

  • Clustering Images
  • Embedding Images: give each images an coord to put similar ones together
  • Embedding Words

Semi-Supervised Learning

Only given a subset of labels

Applications

Label propagation

Self-Supervised Learning

Infer labels from large collections of unlabeled data by exploiting other properties associated w/ dataset
  • Professor Worked on it during PHD

Why Works

Supervised: hitting its limit
Unsupervised: too hard to approach

Applications

Exploiting temporal correlatons of video frames to deduce that image from nearby frames are similar

Reinforcement Learning

Agent finds its own way & use positive/negative reward/punishment for guide

Course Overview

  1. ML Overview: learning & generalization
  1. Probabilistic approach
  1. Linear Parametric models
  1. Non-Linear Parametric models
  1. Unsupervised
 

TODO: Homework 0

Content: questions from probability, calculus, linear algebra (10 questions)
On Brightspace
Sep 13, 2021 00:00