Dynamic Learning: Architectures and Algorithms

Course 4541.676:
Dynamic Learning: Architectures and Algorithms (2010)

School of Computer Science and Engineering, Seoul National University
Instructor: Prof. Byoung-Tak Zhang
TA: Ha-Young Jang (Office : 138-409, Tel : 02-880-5890)
Classroom: 302-209
Time: Mon 11:00-12:15 and Wed 11:00-12:15
Course description:

Conventional machine learning models, including both supervised and unsupervised models, assume a training set of data items, all of which is available when learning starts. However, many real applications involve with dynamic learning situations where the training data come in a stream sequentially and continuously in a long-lasting process. Typically the target distribution of the learning is non-stationary and changing with uncertainty. Conventional learning methods are often insufficient to deal with this. Learning in dynamic environments requires different architectures and algorithms since the objective of learning should, for example, balance the tradeoff between long-term and short-term values. This course reviews the architectures and algorithms developed for dynamic learning settings, and discuss some application examples. A project will be assigned in which the students will develop a dynamical system that learns from video streams (i.e. sequences of picture-text pairs) to perform a visual storytelling.

Evaluation:

One exam (40%)
One miniprojects + Two reports (50%)
Participation in discussion (10%)

Project :

Visual storytelling : Learning to regenerate the sequence of image-text pairs (from video)
- Description Data Notice! Data is changed!(09.17)
References

Lecture Schedule

Date	Topics	Slides
9/1 Wed	Course overview
9/6 Mon	Motivating examples Dynamical systems
9/8 Wed	Dynamic learning Visual storytelling project	Ch1_Intro
9/13 Mon	Autoregressive models Time-series analysis AR, MA, ARMA Term project announced (First report due 10/13)	Ch2_AR
9/15 Wed	Kalman filters Model, algorithms, variants	Ch3_KF
9/20 Mon	No class (Thanksgiving holiday)
9/22 Wed	No class (Thanksgiving holiday)
9/27 Mon	Monte Carlo filtering Monte Carlo approximation Importance Sampling (IR)	Ch4_PF Supplement 1 Supplement 2
9/29 Wed	Particle filters Sequential Monte Carlo (SMC) Vidio lecture : http://videolectures.net/mlss09uk_godsill_pf/
10/4 Mon	Evolutionary Monte Carlo EC, PSO, evolutionary MCMC, EDA	Ch5_EMC Supplement 1 Supplement 2 Supplement 3
10/6 Wed	Monte Carlo and the mind Video lecture: http://videolectures.net/mlss2010_griffiths_mcm/
10/11 Mon	Markov decision processes (MDPs) Definition, problem, solution Partially observable MDPs	Ch6_MDP_RL
10/13 Wed	Reinforcement learning Q-learning Due of project report 1 is postponed for a week.	Ch6_MDP_RL
10/18 Mon	Video lecture: Reinforcement learning
10/20 Wed	Video lecture: Graphical models Due of porject report 1
10/25 Mon	Hidden Markov models (HMMS) Model, learning, inference	Ch7_HMM
10/27 Wed	Conditional random fields (CRFs) Model, learning, inference	Ch8_CRF
11/1 Mon	Dynamic Bayesian networks Architecture, learning, inference Video lecture: Dynamic factor graphs (DFGs)	Ch9_DBN
11/3 Wed	Dynamic hypernetworks Architecture, learning, inference	Ch10_DHN
11/8 Mon	Discussion
11/10 Wed	Discussion
11/15 Mon	No Class
11/17 Wed	Exam
11/22 Mon	Project presentation 1 Bado Lee Eun-Sol Kim and Myung-Goo Kang Byoung-Kwon Lim
11/24 Wed	Project presentation 2 Taemin Park, Joonso Lee and Dooyoung Kim Minkyu Kim and Weerayot Aramphianlert
11/29 Mon	Project presentation 3
12/1 Wed	Future of dynamic learning
12/6 Mon	Due of project report 2

This page is maintained by Eun Sol Kim
Last updated: Sep 27, 2010.