|
COS 402 - Fall 2017
|
|
This
course will survey the aspects of intelligence exhibited in biological systems
and algorithmic approaches to mimic it. Material will include theoretical and
applicative treatment of inductive learning, reinforcement learning, artificial
neural networks, natural language processing and knowledge representation.
In
comparison to previous years - we will cover new topics (natural language
processing, optimization, and deep learning), and will require more
mathematical background for topics we cover.
Oct 5th 19:30 -
garden theater, showing of --Ex-Machina--, Q&A + discussion with Profs.
Hasson (PNI) & Hazan on AI
(Enrichment activity;
attendance is encouraged but not mandatory.)
Number |
Topic / hype title |
Reading:
class material |
Optional Reading |
Problem sets |
Date of class |
Lecturer |
|
Learning from
examples, induction
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
State
of the art |
Norvig-Chomesky
debate
|
|
9/15 |
Hazan |
|
2 |
Decision
trees, symbolic computation (medical diagnosis) Max
information gain rule How
to build decision trees, Overfitting |
Slides lec 2 |
Book
5 chapters 18.2 |
9/20 |
Hazan |
|
3 |
Statistical
learning, sample complexity,
|
Slides lec 3 |
Book
5 chapters 1-3 |
|
9/22 |
Hazan |
4 |
Learning
via efficient optimization. Sample
complexity of python programs -> optimization SVMs
/ linear classification The
perceptron, margin, proof of convergence |
Slides lec 4 |
Book
5 chapter 4.2 |
9/27 |
Hazan |
|
5 |
More
on optimization: mathematical
optimization, convexity, intro to convex analysis, example: learning SVM with SGD
|
Book
6 chapters 2.1,3.1 Book
5 chapter 14.1 |
Book
6 chapter 2 Book
5 chapter 14 Book
7, chapters 3.1-3.2 |
9/29 |
Hazan |
|
6 |
Stochastic
optimization: Stochastic
estimation of the gradient, stochastic gradient descent |
Book
6 chapter 3.4 Book
5 chapter 14.3 |
|
|
10/4 |
Hazan |
7 |
Intro
to Deep Learning: Deep
nets. Non-convex optimization. Training via backpropagation algorithm. |
|
|
10/6 |
Arora |
|
8 |
Vision
via Deep Nets: Neural nets for image recognition. Convolutional architectures. Deep nets. Regularization strategies. |
Draft of book on Deep Learning by
Goodfellow, Bengio, Courville. |
Recognizing Digits. (Due Oct 18) |
10/11 |
Arora |
|
|
|
|
|
|
|
|
|
Language &
communication, NLP
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
Language
Models: Intro Introduction+history Markovian language models. n-grams. Smoothing |
10/13 |
Arora |
|||
10 |
Word
embeddings Word2vec,
PMI, etc. Using word embeddings to solve analogies
and other tasks. |
Optional
readings 1.
by T. Landauer and S. Dumais , 2. by P.
Turney and P. pantel , 3. by Sanjeev
Arora |
10/18 |
Arora |
||
11 |
Collaborative
filtering, latent factor models. |
|
|
10/20 |
Hazan |
|
12 |
Logic: Background
on logic, Formal
definitions, derivation, verification |
R&N (Book 8) Section 7.3,
7.4. Skim through 7.5.1 and 7.5.2 (related to what we did, but more detailed) |
10/25 |
Arora |
||
13 |
Midterm |
|
|
10/27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Knowledge representation
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
Knowledge
Representation and Reasoning: Bayesian
Networks Marginalization
|
Chapter 2
from Bayesian
Artificial Intelligence; A survey due to
Kevin Murphy |
Movie
Embedding Due Nov. 15th |
11/8 |
Arora |
|
15 |
Bayes
nets: Definition
of probabilistic Bayesian nets Modeling
via Bayes nets Inferences |
Bayesian
Networks witout tears by Eugene Chamiak |
. |
11/10 |
Arora |
|
16 |
Markov
Chain Monte Carlo: The
sampling problem, simple sampling methods, Markov
chains, stationarity, ergodicity The
MCMC algorithm |
Lectures
12 and 13 from the course stochastic
simulation. |
|
11/15 |
Hazan |
|
17 |
Hidden
Markov Models Temporal
models, application to text tagging Viterbi
decoding algorithms |
Lecture
notes by Percy
Liang And
by Michael
Collins |
Bayesian
networks, due Nov 29 |
11/17 |
Hazan |
|
|
|
|
|
|
|
|
|
Reinforcement
learning
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
Game
playing Search,
A^* heuristic |
|
|
11/22 |
Arora |
|
19 |
Reinforcement
learning, MDP: Define
RL, MDP Markov
chains, Markov Reward processes, Ergodic
theory reminder, Bellman equation |
Book by
Sutton and Barto |
11/29 |
Hazan |
||
20 |
Dynamic
programming The
Bellman equation Policy
iteration |
|
12/1 |
Hazan |
||
21 |
Algorithm
for RL Q-learning,
function approximation |
|
12/6 |
Hazan |
||
22 |
Guest
lecture, deep learning |
|
|
|
12/8 |
Seung |
23 |
Exploration
- MAB problem UCB,
EXP3 algorithms & analysis |
Book by
Sutton and Barto, ch2 |
|
12/13 |
Li |
|
24 |
Ask
us anything |
|
|
|
12/15 |
Arora+Hazan |
|
|
|
|
We'll
post additional optional reading as the course progresses.
Motivation: Some discussion and collaboration
enhances your educational experience, but too much collaboration---in the
extreme case, copying each other's solutions--- is unethical and detrimental,
and also leave you ill-prepared for the exams, which count for 50% of the
grade.
? OK:
discussion with others about the material in this course, including HW (if
attempted alone first), in which case names of all discussants should be noted
for each problem separately. Comparing and discussing the results of
experiments.
? NOT OK:
No copying of any sort from any student (past,
present or future), from the web, from prior year solutions, from any other
course or source. Do not take notes on any solution that may have been arrived
at during discussion; instead try to reconstruct it independently later when
you write it down. Consulting any kind of website, blog,
forum, mailing list or other service for discussing, purchasing, downloading or
otherwise obtaining or sharing homework solutions is strictly forbidden (except
for piazza and the class mailing list).
Deviations from this policy will result in
university disciplinary action.
Exercises: 10% off for every day (24
hours) late. Exercises not accepted more than 4 days late.
Special circumstances: please send letter from residential college dean.