• login

CS295, Fall 2008: Research Problems in Machine Learning

Instructors: Alex Ihler and Max Welling

Class schedule: Mon/Wed 2:00-3:30pm, DBH 1423

During this class you will engage in research in small groups of approximately 3 to 4 students. Each group will select a research project as their focus for the term. A list of several potential projects will be provided by the instructors, or students may define their own research project as well (subject to approval, i.e., that it be sufficiently challenging and not identical to the research you were doing for your PhD anyway). We strongly encourage projects that will either lead to tangible results (e.g. a publication or to a web-based application). See below for some example projects.

For the course, students will be required to read and present to the class papers relevant to their project, and write a research paper or technical report and present the results of their work. Grading will be satisfactory/unsatisfactory.

Students will meet as a class on Mondays, and as individual groups with the instructors on Wednesdays.

Reading and Presentation Schedule

October 6th

• Jon Hutchins -- Jigsaw Puzzle Solver Using Shape and Color -- PDF
• Ronen Vaisenberg -- Finding Interesting Associations without Support Pruning -- PDF

October 13th

• Drew Frank -- Affinity-based clustering -- PDF
• Ian Porteous -- General intro to financial problems

October 20th

• Yutian Chen -- Compression-based texture merging -- PDF
• America Holloway -- Models of influence (e.g. herding, flocking in physics and graphics, cascades, and rumor propagation)

October 27th

• Todd Johnson -- Image segmentation using information theory and curve evolution (pdf)
• Vikram Bodicherla -- Volatility prediction using NEWS items

November 3rd

• Sidharth Shekhar -- Intelligent Stock Trading by Turning Point Confirmation and Probabilistic Reasoning PDF -- Slides
• Vijay Rajakumar--Dynamics of rumor propagation on small-world networksPDF

November 17th

• Kiran Shivaram -- A Global Approach to Automatic Solution of Jigsaw Puzzles PDF

Example (potential) research projects for the class:

Jigsaw puzzle assistant

Imagine a user uploading an photo of the pieces of a jigsaw puzzle. Your task is build a system that can suggest moves to the user, or potentially solve the puzzle. Possible technical ingredients for this example include:

• image segmentation
• contour/shape extraction
• feature extraction (colors, edges, etc.)
• optimization/inference to solve the puzzle
• testing performance.

Stylistic or corrective image transforms

These guys proposed learning transformations from one image to another, which can be used to "impose" a particular image style onto another image. They use it for producing faux artistic effects, and for image restoration such as super-resolution and deblurring (ill-posed) inverse problems.

Automatic identity tagging

Currently, there exist a large number of tools to search and organize photos by date, folder, EXIF tags, etc. However, almost no one bothers to tag images with identity information, because it's too much effort. But what if you could do all that automatically, by detecting faces, grouping them, and getting a relatively small amount of user input instead? Aspects include

• face detection
• feature extraction
• clustering

Google apparently does something like this on PicasaWeb (described here); see also for example this paper.

Image search / retrieval

Because images are typically untagged with any identifying information, it becomes important to find images using some criterion (which has yet to be decided upon). As an example, see this demo for searching for images based on color, or this demo which searches based on wavelet coefficients.

Alternatively, you might want to find images which are "similar" to some test image. For example, they might contain the same entity or object -- given an image of some animal, plant, etc. you might want to know what it is by searching a database of labeled images for something with a similar appearance.

Intelligent web site-map

Automatically create an organization and "map" or hierarchical interface given a web address. For inspiration, see this CalIT2 topic model. One goal here is to make something like this truly automatic, so that no (or almost no) intervention is required. You might use social network analysis, topic models of text, etc. to achieve this.

Other ideas

• Collaborative filtering
• Google earth or google street view
  • Combine with other geospatial data? Photos, economic data, etc?
• Robotics / mapping