SNU Biointelligence Lab

4190.676 Self-learning Neural Algorithms (Fall 2017)

(a.k.a. Artificial Neural Networks)

2017 Fall Graduate Course in Computer Science and Engineering and Brain Science

   
Instructor Prof. Byoung-Tak Zhang
Main TA Jiseob Kim (jkim@bi.snu.ac.kr)
Sub TA Je-Hwan Ryu (jhryu@bi.snu.ac.kr)
Classroom 302-107 (originally 302-209)
Time Tue & Thu, 14:00–15:15

Course Objectives

We study “self-learning” networks, i.e. models that learn in an unsupervised and “self-supervised” way without the help of an explicit teacher. These models are neuro-biologically inspired and, usually, self-organizing, dynamic, recurrent, and auto-encoding networks. We examine the principles of neural learning algorithms from the historical models, such as Willshaw-von der Malsburg feature maps, Linsker models, Kohonen’s self-organizing maps, Grossberg models, recurrent networks, Anderson’s brain-state-in-a-box, actor-critic networks, Hopfield’s associative memory, Boltzmann machines, and deep belief networks.

We study mathematical tools for approximation and optimization of the neural learning models. These include information-theoretic algorithms, such as maximum entropy, mutual information, and KL divergence as well as the statistical-mechanical methods, such as Markov chains, Metropolis algorithms, Gibbs sampling, and simulated annealing. We also examine the neurodynamic models of self-supervised, end-to-end learning to solve the challenging problems, such as time series prediction and reconstruction. These include Markov decision processes, approximate dynamic programming, reinforcement learning, sequential Bayesian estimation, Kalman filtering, particle filtering, real-time recurrent learning, dynamic reconstruction of a chaotic process.

Textbooks

Evaluation

Component Weight
Two exams (midterm + final) 80%
Homework 10%
Participation and discussion 10%

Lecture Schedule

Week Dates Topics Slides
Week 1 9/5, 9/7 Learning in Neurodynamic Self-organizing Systems — Neural Networks, Unsupervised / Self-supervised Learning; Mathematics for Neural Learning; Principal-Components Analysis (Ch. 8): PCA, Hebbian-Based Maximum Eigenfilter, Hebbian-Based PCA, Generalized Hebbian Algorithm, Kernel PCA Ch8_PCA
Week 2 9/12, 9/14 Self-organizing Maps (Ch. 9) — Willshaw-von der Malsburg Model; Kohonen’s SOM Model Ch9_SOM
Week 3 9/19, 9/21 Information-Theoretic Learning Models (Ch. 10) — Maximum Entropy, Kullback-Leibler Divergence; Mutual Information, Infomax, ICA Ch10_ITLM
Week 4 9/26, 9/28 No Class — Statistical-Mechanical Learning Methods (Ch. 11): Statistical Mechanics, Markov Chains; Metropolis, Gibbs Sampling, Simulated Annealing Ch11_SM
Week 5 10/3, 10/5 Korean Thanksgiving Holiday  
Week 6 10/10, 10/12 Makeup Class (10/12 at 19:30) — Deep Neural Networks (Ch. 11): Boltzmann Machines; Deep Belief Networks Ch11_SM
Week 7 10/17, 10/19 Dynamic Programming (Ch. 12) (10/17); Problem Solving Session by TA (10/19) Ch12_DP
Week 8 10/24, 10/26 Summary (10/24); Mid-term Exam (10/26)  
Week 9 10/31, 11/2 Dynamic Programming (Ch. 12) — Markov Decision Process, DP, Bellman Equation; ADP, Reinforcement Learning, TD, Q; (11/2) PACS-2017 conference (see announcements) Ch12_DP
Week 10 11/7, 11/9 Neurodynamic Models (Ch. 13) — Dynamic Systems, Attractors, Chaos; Hopfield Models, Dynamic Reconstruction Ch13_ND
Week 11 11/14, 11/16 Classroom Change (Mogam Hall, Bldg 500 on 11/16) — Bayesian Filtering (Ch. 14): State Space Models; Kalman Filters, EKF, CKF Ch14_BF
Week 12 11/21, 11/23 Makeup Class (11/21 at 19:00) — Particle Filters (Ch. 14): Approximate Bayesian Filtering, Particle Filters, SIR Algorithm; Dynamic Recurrent Networks (Ch. 15): Recurrent Network Architectures, Backpropagation through Time Ch14_Suppl / Ch15_RN
Week 13 11/28, 11/30 Real-Time Recurrent Learning (Ch. 15) — RTRL Algorithm, Vanishing Gradients; EKF Algorithm for Training RMLP; Final Exam (11/30) Ch15_RN
Week 14 12/5, 12/7 Review and discussion  

Announcements