March 2011 Archives

Audio Mostly call for papers

| No Comments
The conference Audio Mostly "A conference on interaction with sound" will this year be held September, 7 - 9 in Coimbra, Portugal:

Audio in all its forms - music, sound effects, or dialogue - holds tremendous potential to engage, convey narrative, inform, create attention and enthrall. However, in computer-based environments, for example games and virtual environments, the ability to interact through and with sound are still today underused. The Audio Mostly Conference provides a venue to explore and promote this untapped potential of audio by bringing together audio experts, content creators and designers, interaction designers, and behavioral researchers.

The area of interest includes interactivity through sound, tools and methods to support sound design work and evaluation and new and innovative applications of sound. It implies cognitive, psychological and social research studies, as well as applied research and technological innovations in audio analysis, processing and rendering. The aim is to both describe and push the boundaries of sound-based interaction in various domains, such as entertainment, education, safety and health care.

The theme this year is "Sound and Context"
More information can be found at:

Important dates:

Deadline for submission - Friday, April 29
Notification of acceptances - Friday, June 3
Final paper camera ready - Friday, July 8
Deadline for author registration - Friday, July 8
Conference - September, 7 - 9
Following on the heels of the previous post, and the post previous to that, post previous to that, and the post previous to that, I am now comparing recovery performance of Probabilistic OMP (PrOMP), and straight-up OMP. In my implementation of PrOMP, I limit the number of independent trials to 10. Furthermore, I make \(\epsilon = 0.001\), and I set the number of high probability atoms to 2. The end condition tolerance is the same as for OMP (\(||\vr||/||\vx|| < 10^{-5}\)). Click on the image below to see the results.
Following on the heels of the previous post, post previous to that, and the post previous to that, I am now comparing recovery performance of \(\ell_1\) minimization, and cyclic MP. Click on the image below to see the results.
Following on the heels of the previous post, and the post previous to that, I am now comparing recovery performance of the recommended version (by Maleki and Donoho) of two stage thresholding (TST) --- which is essentially subspace pursuit without the oracle --- and cyclic MP. Click on the image below to see the results.

So IHT does well after all!

| No Comments
Following on the heels of the previous post, I am now comparing recovery performance of the recommended versions (by Maleki and Donoho) of iterative hard thresholding (IHT), and iterative soft thresholding (IST). Click on the image below to see the results.
The other day, I posted some initial results of cyclic matching pursuit (CMP) for compressed sensing recovery. This morning, I came to an office full of new test results! If you click on the image below, you will see some of these!

Why blog research?

| 1 Comment
When I started this research blog over one year ago (with this non-triumphal post), I began my own personal experiment in how I do research. And 189 entries later, I am quite happy with the results and looking forward to doing it as long as it keeps working.
This news just in: Joan Serrà of the fantastically productive and influential Music Technology Group at the Universitat Pompeu Fabra, Barcelona, has just successfully defended his dissertation on an excellent topic: "Identification of versions of the same musical composition by processing audio descriptions". I have discussed his work here before in relation to cover songs, for example, here and here. I am looking forward the most to learning about (and using) the non-linear time series analysis techniques he uses! This is excellent work.
I have discussed Cyclic Matching Pursuit here (original paper) and here (extension to time-frequency dictionaries, and an orthogonal least squares flavor) before. With reference to the recovery of compressively sampled sparse signals, I have presented some results here and here and here. I have shown that MP, the "pure greedy algorithm" itself, performs horribly for recovery. Spending some time orthogonalizing the residual, as in OMP, really helps things along.

5 days

Yesterday, I distributed my simulations between two computers (which involved me spending about 2 hours troubleshooting Windows, e.g., "Where is the file I just downloaded?", and "Why on Earth doesn't Windows come with a program to mount .iso files?", and "I need administrator privileges? But it says I have administrator privileges!", and "Oh, I need to log in locally using "Computer-18983\bob" and not just "bob"?", and "How the hell do I type a "\" on a Danish Windows keyboard?"). To avoid storing the several dozen gigabytes of randomly generated sparse signals and sensing matrices that is my "problem suite", I decided instead to use the same seeds on both computers, which took some time as well since I tried to use RandStream. I just ended up using the legacy approach with "rand(k,seed)". And finally, I adjusted my simulations with the following changes:

  1. reduced problem dimensionality by half, from 800 (Maleki and Donoho) to 400 (but still keep almost 900 pairs of sparsity and indeterminacy values, and 100 trials with each)
  2. once the recoverability success of an algorithm reaches 0, stop running that algorithm (!)
  3. Do not even test recoverability when the indeterminacy is 1. That means no compression, so who cares?
  4. only run OMP, BP, IHTR, ISTR, and TSTR.
With all of these changes, I am only five days from writing the paper from which I was 143 days yesterday. What a difference a day makes! (My anger with Windows will take some more time to recede.)

143 days

| No Comments
That is how long I must wait for my 5400 simulations to finish running. I started this process more than 50 hours ago, thinking it would be done Tuesday. Maleki and Donoho are not kidding when they write,
It would have required several years to complete our study on a single modern desktop computer.
In the Maleki and Donoho paper, they generalize CoSaMP (D. Needell and J. A. Tropp, "CoSaMP: Iterative signal recovery from incomplete and inaccurate samples," Appl. Comput. Harmonic Anal., vol. 26, no. 3, pp. 301-321, 2008), and subspace pursuit (W. Dai and O. Milenkovic, "Subspace pursuit for compressive sensing signal reconstruction," IEEE Trans. Inf. Theory, vol. 55, no. 5, pp. 2230-2249, May 2009), as "composite iterative algorithms," which they call "two-stage thresholding" (TST).

Today I reconciled my code with Blumensath's, and am now confident I have an honest implementation of iterative hard thresholding (IHT). My initial surprise at how poorly IHT was performing for all but the embarrassingly sparsest signals appears to hold. For guidance, I have revisited the optimal tuning work by Maleki and Donoho. They rightly observe in this paper that we are awash in recovery algorithms that are not ready-to-run ... well, maybe not awash, but I sure feel like it. And with IHT and its two parameters, and little guidance for how to set them, I was beginning to question its utility.
I received the greatest piece of fan mail I have ever received a few days ago:

Hello! I have been a big fan of 50 Particles in a Three-Dimensional Harmonic Potential for almost 10 years now. ... Fast forward to now, and here is why I'm emailing you. Tomorrow night I am DJing a gig in Melbourne, Australia, and I'd really like to open my set with 50 Particles, or at least a portion of it. ... I will be playing to a crowd of about 200 people on a very powerful, very hi fidelity sound system. ... The rest of my set will be a kind of dance music called psytrance, but I really want to open my set with something that people will remember. ... Is there anywhere that a CD quality version of 50 Particles already exists on the net that you know of?
Since I never imagined how 50 Particles could ever fit as an introduction for a DJ at a rave, I just had to satisfy that request!

And startled I am by the awesome results! He timed my Coulomb explosion perfectly! DJ Sunkid spinning at the Big Red Bus (if you arrive on a bike you get in free!) March 2011:

This really makes me want to get out my huge Fuel pants, wallet chain, silver sneakers, and umpire shirt! PLUR.
Over the past few days, I have been running some experiments with iterative hard thresholding (IHT). I have described this approach to sparse signal recovery from compressive measurements here. This method is presented and analyzed in T. Blumensath, M. E. Davies, "Iterative Hard Thresholding for Compressed Sensing," Applied and Computational Harmonic Analysis, vol. 27, no. 3, pp. 265-274, 2009.
The problems associated with Matching Pursuit using an overcomplete dictionary (the pure greedy algorithm in approximation theory) have long been known: poor convergence, and bad choices. I provide a short synopsis below of work that I know of within the discipline of signal processing that attempts to deal with these problems; but this work is relatively audio-centered, and I am sure there exists work in the image processing domain.
Multimodal interaction in virtual environments
Aalborg University Copenhagen
May 10,11,12, 13 and June 6, 2011

This 4-ECTS course provides an overview of multimodal interaction techniques for virtual environments. We start with an overview of multimodal perception to explain how humans behave in virtual environments where incomplete and impoverished sensory cues are reproduced. We then present an overview of technologies for visual-haptic-audio feedback in virtual environments, together with sensing technologies based on capacitive sensing and optical motion capture. We discuss issues of integration of technologies, and we describe algorithms for recognizing input data as well as simulating feedback based on physics modelling. We then introduce evaluation techniques for multimodal environments. 

The course is 4 ECTS, divided into 2 ECTS of lectures and 2 ECTS of mini-project. 

 Introduction to multimodal interaction in virtual environments: perceptual illusions, sensory substitution, and multimodal enhancement.
-Visual feedback: screen, projectors, head-mounted display
-Auditory feedback: surround sound, headphones
-Technologies for haptic feedback
-Physics based algorithms for audio-haptic feedback
-Sensing and tracking technologies (capacitive sensors, optical motion capture)
-Integration of technologies
.Evaluation of multimodal interfaces

If you wish to attend, please register BEFORE 15 APRIL on the following page:

choosing the registration form for the course: MULTIMODAL INTERACTION IN VIRTUAL ENVIRONMENTS

For more information please contact:
Associate professor Stefania Serafin
Aalborg University Copenhagen

Blog Roll

About this Archive

This page is an archive of entries from March 2011 listed from newest to oldest.

February 2011 is the previous archive.

April 2011 is the next archive.

Find recent content on the main index or look in the archives to find all content.