00intro 1
00intro 1
TAs:
Su Yeong Lee (CAAM)
Kexiang Wang (CAAM)
2
2/43
/43
Topics
1. Clustering
2. Dimensionality reduction
3. Manifold learning
4. Regression
5. Online algorithms
6. Kernel methods (Hilbert space algorithms)
7. Bayesian learning
8. Deep learning
9. Generative models
3
3/43
/43
Prerequisites
4
4/43
/43
Support
Recitations:
• On an as needed basis, place and time TBD
Office Hours:
• Fridays 1pm Crerar 221
Online:
• canvas.uchicago.edu (slides, lecture notes, assignments and grades)
5
5/43
/43
Resources
Books (Strictly optional! More for “further reading” than anything else.)
• Kevin Murphy: Machine Learning: A probabilistic perspective (2012)
Warning: very Bayesian
• Zhang, Lipton, Li and Smola: Dive into deep learning (d2l.ai)
• Hastie, Tibshirani, Friedman: The Elements of Statistical Learning (2008)
(available electronically on the library’s web site)
Online Courses
• Andrew White’s book “Deep learning for molecules and materials”
https://dmol.pub/index.html
6
6/43
/43
Credit
7
7/43
/43
8
8/43
/43
What is Machine Learning?
Two types of programming
10
/43
10/43
11
11/43
/43
Machine Learning in the abstract
f : x 7→ y
12
12/43
/43
Nomenclature
13
13/43
/43
Deductive vs. inductive inference
• Deductive inference:
rules −→ data
• Inductive inference:
data −→ rules
14
14/43
/43
Typical ML task 1: Regression
15
15/43
/43
Typical ML task 2: Classification
16
16/43
/43
Typical ML task 3: Ranking
Internet search
Elections Sports
17
17/43
/43
Typical ML task 4: Clustering
18
18/43
/43
ML task 5: Dimensionality Reduction
19
/43
19/43
Applied vs. theoretical ML
This course will focus on the fundamental algorithms rather than specific
applications.
20
/43
20/43
Origins: Classical Artificial Intelligence
AI vs. ML
22
22/43
/43
Early attempts
23
23/43
/43
Formal reasoning = intelligence?
24
24/43
/43
Is the brain just a computer?
Pitts & McCullogh show that neurons appear to perform simple logical
operations (1943)
“So if all that the brain does is such mechanistic operations, then it should be
easy to imitate on Turing machines (i.e., computers)”
25
/43
25/43
The Turing test
In his landmark 1950 paper “Computing Machinery and Intelligence” Turing
proposes a positivist approach: “If a machine can fool a human into thinking
that it is a human, then it must be intelligent” → Weak AI
Prediction: “By the year 2000 machines with 120MB of memory would be
able to fool 30% of human judges in a 5min test”.
26
26/43
/43
Objections to the Turing test
Even if a computer passes the Turing test it cannot be truly intelligent
because...
1. Theological: computers have no soul
2. “Head in the sand”: it would be too scary
3. Mathematical: Godel incompleteness and such
4. Consciousness: Searle’s Chinese room argument
5. Disabilities: a machine will never be able to do fall in love/invent jokes/tell
right from wrong/etc.
6. Lady Lovelace’s: will never do anything original
7. The brain is not digital
8. The brain is not predictable
9. Extra-sensory perception
27
27/43
/43
28
28/43
/43
The Dartmouth conference (1956)
29
29/43
/43
True beginnings: from philosophy to
building things
“We propose that a 2 month, 10 man study of artificial intelligence be carried
out during the summer of 1956 at Dartmouth College in Hanover, New
Hampshire. The study is to proceed on the basis of the conjecture that every
aspect of learning or any other feature of intelligence can in principle be so
precisely described that a machine can be made to simulate it. An attempt
will be made to find how to make machines use language, form abstractions
and concepts, solve kinds of problems now reserved for humans, and
improve themselves. We think that a significant advance can be made in one
or more of these problems if a carefully selected group of scientists work on it
together for a summer.”
30
/43
30/43
Early successes
31
31/43
/43
AI winters ’74-’80, ’87-’93
32
32/43
/43
New beginnings: Machine Learning
The birth of Machine Learning
34
/43
34/43
The old vs. the new AI
New: pragmatic, focused on specific tasks, much closer ties to math and
statistics than neuroscience and logic, driver behind lots of technologies
Question: Classically, the subject that deals with the art of learning from data
is Statistics. So is ML just a branch of Statistics? No.
35
/43
35/43
Statistics
Nonaparametric statistics
Bayesian statistics
Probability
Empirical Process Theory
Computer Science
Artificial Intelligence
Computational Learning Th
Complexity Theory
Randomized Algorithms
Machine Databases
Learning Distributed Systems
Mathematics
Functional Analysis
Random geometry
Optimization
Numerical analysis
36
36/43
/43
Applications
NLP
Speech recogni-
tion
Translation
Computer Vision Summarization
Object detection Grading
Object recognition Search & rec.
Structure from motion Web search
Collaborative filtering
Ad placement
etc., etc.
Machine
Robotics
Learning
Autonomous vehicles
Robot assistants
Medical
Detection & imaging Finance
Automated diagno- High freq. trading
sis Portfolio selec-
tion
Comp Bio Risk analysis
Protein structure
Systems bio
37
37/43
/43
Hallmarks of ML
ML is ambitious:
• Datasets are often very high dimensional (∼ O(105 )) .
• Data is often abstract (structured objects vs. just vectors).
• Datasets are massive (∼ O(108 ) examples ) .
• Really want to build actual systems that work.
ML is brutal:
• Don’t need to think hard about the domain because with enough data,
even black box algorithms work really well (really?).
• Butcher the statistics as much as necessary to get an algorithm which
actually runs.
• Insist on algorithms that run in time
O(m3 ) → O(m2 ) → O(m) → o(m) .
38
38/43
/43
Taxonomy of Machine Learning
Taxonomy of machine learning 1.
40
40/43
/43
Taxonomy of machine learning 2.
41
41/43
/43
Taxonomy of machine learning 3.
• Batch learning: see whole training set first, then predict on test
examples.
• Online learning: examples are presented one-by-one, first try and
predict yt , then find out what yt really is and learn from it.
• Transductive learning: like batch, but know test x′i ’s at training time.
• Active learning: algorithm can ask for next data point
• Reinforcement learning: exploring the world incurs a cost (games,
robotic control)
42
42/43
/43
Taxonomy of machine learning 4.
43
43/43
/43