CS680: Special Topics in Computer Science

Advances in Image Understanding and Machine Learning

Spring 2014

 

Berrin Yanıkoğlu, Hakan Erdoğan

 

 

Course Material

 

 

Brief Description

 

The objective in this course is to follow recent developments in computer vision and machine learning; especially relating to object detection and recognition, image retrieval and related techniques in machine learning. In particular, we will look into the topic of deep learning in depth, as well as topic models, word embeddings, new object representations, and possibly structured, online, semi-supervised learning methods etc. according to interest.

 

We are going to read scientific papers (about 4-6 per week) at varying depth (learn state-of-the-art over some application papers versus understand this method in sufficient detail) and follow tutorials (about 3-4 tutorials during the course).

 

The class will be participatory with students and faculty members presenting/leading papers/tutorials in turn. Each student will also be expected to implement a new technique on a simple/toy problem (e.g. deep learning on a simple problem), as well as working towards on one of the assigned current problems (e.g. in parallel with a large scale object recognition or retrieval competition).

 

There will be no exams, grading will be done according to class participation/contribution (60%) and homework/project (40%), with roughly the given weights.

 

Prerequisite: Machine Learning (CS516) and/or Pattern Recognition (EE566).

 

Intended Audience: Advanced graduate students (2nd year MS or PHD students).

 

Tentative Grading: Course participation (attendance and contribution): 20%    Presentations: 40%     Hws/Project:40%.

 

Course participation is measured through students’ involvement in the discussion about the presented material.

Presentation grade will be given for student  presentation of a chosen paper describing some state-of-the-art approach.

Project will involve some implementation and/or improvement of an existing research. It can be done in groups of 2, and

Ideally it would result in a conference paper or a technical report.

 

 

Schedule and Reading Materials:

 

We will start with http://ufldl.stanford.edu/wiki/index.php/UFLDL_Tutorial

 

The tutorial presentation will be mixed with presentations of papers describing state-of-the-art work in large scale object recognition and in particular those using  deep learning. For instance http://www.image-net.org/challenges/LSVRC/2013/iccv2013.php

 

At least the initial few presentations will be mostly at overview level.

 

http://deeplearning.net/reading-list/tutorials/