machine learning algorithmic aspects ofhuman learning, machine learning take a course understand the...
TRANSCRIPT
![Page 1: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/1.jpg)
Algorithmic Aspects of Machine Learning
COMPSCI 590.7 Fall 2015Rong Ge
![Page 2: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/2.jpg)
Basic Info
● Webpage (course materials, related papers): http://www.cs.duke.edu/courses/fall15/compsci590.7/
● Contact me:LSRC D226Email: [email protected]
● Email List:You are already on the list if you registered.Send me an email otherwise.
![Page 3: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/3.jpg)
Basic Info
● Expect to see (and do) Proofs.● Homework:
~3 problem sets, due 2 weeks after posted in class.Discussions allowed but must acknowledge.
● Final Project:Form groups of size 2-3, ~1 month.Can be original research or literature survey.
![Page 4: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/4.jpg)
Lecture 1:Machine Learning Basics
![Page 5: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/5.jpg)
Machine Learning Basics
● What do we want “machines” to learn?
● How do we know machines have learned?
● What guarantees can we hope to get?
● What tools do we have?
![Page 6: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/6.jpg)
What do we want machines to learn?
![Page 7: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/7.jpg)
Example: Dogs vs Cats
https://www.kaggle.com/c/dogs-vs-cats
Input: Many images with labels (cat or dog)Goal: Given new image, decide cat or dog.
![Page 8: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/8.jpg)
Example: Netflix
Input: Movie ratings from usersGoal: Recommend new movies for users
![Page 9: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/9.jpg)
Example: Community Finding
Input: List of friendsGoal: Find “communities”
![Page 10: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/10.jpg)
Supervised vs Unsupervised
● Input: (data, label)● Output: Function f● Hope:f(data) = label
Supervised
● Input: data● Output: “structure”● Hope: explain and predict
Unsupervised
![Page 11: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/11.jpg)
Example of Structure
Movies
Users Users
MoviesGenre
Genre
● Movies have different genres.● Users like different genres.● Learning ~ Find the genres of movies and users● To recommend
○ determine what genres the user likes○ recommend a good movie in that genre.
![Page 12: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/12.jpg)
Example of Structure: Probabilistic Models
● People are in different communities● Same community ⇒ Higher probability p to know each other● Different community ⇒ Lower probability q to know each other● Learning ~ Find out which groups of people are in the same community
![Page 13: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/13.jpg)
Probablistic Models
Given: Data
Assumption: Data is generatedfrom a class of distributions(“model”)
Learning ~ Findingparameters of the model
![Page 14: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/14.jpg)
Connection
MNIST: Given imagesrecognize digits.
Unsupervised: Given images,cluster similar images
Unsupervised Learning ⇒ features for Supervised Learning
![Page 15: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/15.jpg)
How do we know machines have learned?
(General approach: Give an exam)
![Page 16: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/16.jpg)
Human Learning, Machine Learning
● Take a course● Understand the course
material● Take an exam
Goal of Exam: If the students understood the material, do well in exam.
● Get training examples● Learn a hypothesis
- maps input to labels● Test on new examples
Goal of Test: If the hypothesis is good, do well in test samples.
![Page 17: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/17.jpg)
What is a good hypothesis?
Ways to fail an exam● Do not understand examples given in class
(Make many mistakes on training samples)● Only memorize the basic examples
(Come up with a complicated hypothesis)
Good hypothesis = Simple + Do well on training
![Page 18: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/18.jpg)
PAC Learning
[Valiant 1984] For a concept class C, for any distribution D on data points, given samples
(x,y): x~D, y = c(x) for fixed c in Can algorithm PAC learns the concept class if with probability 1-δ the algorithm outputs a function f such that
Prx~D[f(x) = c(x)] ≥ 1 - ϵ
![Page 19: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/19.jpg)
● Get samples and labels
● f may not be a box
Example
● Concept Class: Boxes-
+ +
++
+
+++
--
-
- -
-
-
-
+
+
+
-
-
-
● Tested on new samples
GeneralizationIf f is “simple”, doing well on the current samples guarantees good performance on future samples.
![Page 20: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/20.jpg)
Exams for Unsupervised Learning
Make (probabilistic) predictions
What other movies do this user like?
What other people do this user know?
GeneralizationIf the model is “simple”, explaining the current samples guarantees good prediction on future samples.
![Page 21: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/21.jpg)
Alternative Guarantees
● Parameter Estimation:○ Estimate parameters of the model
- (what movies are comedies, who are in same community…)
○ Assumes the model is “correct”.○ Easier to work with, often see in
this course● Maximum Likelihood Estimation:
○ Find the parameters that are best at explaining data○ Only hope to do as well as the “best model”○ Still restrict to the model (like restricting f ~ box)
All models are wrong, but some are useful.
![Page 22: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/22.jpg)
Learning Parameters vs Making Predictions
Parameter Estimation
Making Prediction
Often
Not Necessary
● Has a well-defined model.
● Easy to explain/interpret.● Concise representation.
● More robust to model mismatch
● More efficient algorithms.● Often what we really want
![Page 23: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/23.jpg)
What guarantees can we hope to get?(and why do we care about guarantees?)
![Page 24: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/24.jpg)
Why do we want guarantees?
● Understand why learning works (or not work)○ Does the algorithm work with high dimension/high
noise/…?● Make sound conclusions
○ “I tried to find a good set of parameters but failed”○ vs. “If there is a good set of parameters I’ll find it”
● Design better algorithms○ Can I tweak the algorithm if data is not ideal?○ Get new ideas by different way of thinking.
![Page 25: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/25.jpg)
What we are not likely to do
● PAC learning for many problems○ intersection of halfplanes○ 2 layer “neural network”
● Learn many functions even with specific distribution○ Learning parities with noise...
● Maximum Likelihood for many problems○ mixture of Gaussians,○ topic models
![Page 26: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/26.jpg)
The Hope
Machine Learning Problem (worst case, for every...)
Natural Instance
Reasonable Model Reasonable Assumptions
![Page 27: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/27.jpg)
Example: Netflix
Input: Movie ratings from usersGoal: Recommend new movies for users
![Page 28: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/28.jpg)
Natural Instances
● Models are not perfect, but they can work○ Breaking movies into genres is not perfect, but
allows reasonable commendation
● Natural instances are often easier○ People usually have a preference over different
genres, ...
● “What is natural” is the hard problem○ We will see examples later in the course.
![Page 29: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/29.jpg)
What we will see
M=
A
W
Matrix Factorization
Social Network
Topic Models
Independent Component Analysis
Tensor
Tensor Methods
![Page 30: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/30.jpg)
What tools do we have?
![Page 31: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/31.jpg)
Tools
● Geometry○ Nonnegative constraints○ Finding extreme points
● Linear Algebra○ Spectral methods○ Tensor decomposition
● Optimization○ use convex programs
(LP, SDP) in learning○ optimize a nonconvex function
● Hope: Will have more tools!
Tensor
![Page 32: Machine Learning Algorithmic Aspects ofHuman Learning, Machine Learning Take a course Understand the course material Take an exam Goal of Exam: If the students understood the material,](https://reader030.vdocument.in/reader030/viewer/2022041021/5ed0f8832b6d4e0fbe17d489/html5/thumbnails/32.jpg)
Schedule
● Nonnegative Matrix Factorization and TopicModels (~4 lectures)
● Spectral Clustering (~2 lectures)
● Tensor Decompositions (~5 lectures)
● (Non-convex) Optimization (~5 lectures)
● Matrix Completion (~4 lectures)