Tuesday, July 21, 2009

CS: A postdoc at Simon Fraser University, Dictionary-based Methods (Sparse Approximation) and Audio Signals, IJCAI-09

Jie Liang just sent me the following announcement for a Postdoc position at Simon Fraser University, Vancouver, Canada. Details: A joint postdoc position is available immediately in the School of Computing Science and the School of Engineering Science, Simon Fraser University, Burnaby (metro-Vancouver area), British Columbia, Canada. Candidates should have a PhD degree (or be near completion) in a relevant field and a strong research and publication record. Areas of interest include overlay/peer-to-peer networking, wireless sensor networking, social networking, image/video coding, cooperative/joint source channel coding, distributed source coding, and compressive sensing.

The duration of the position is 12 months and may be extended an additional 12 months.

Please send your full CV in PDF format to Dr. Jiangchuan Liu (jcliu@cs.sfu.ca) and Dr. Jie Liang (jiel@sfu.ca).

Additional information can be found at


This announcement will be featured shortly on the Compressive Sensing Jobs page.



Bob Sturm will be giving a tutorial on sparse approximation and audio data at the 6th Sound and Music Computing conference in Portugal. The title of the talk is Dictionary-based Methods (Sparse Approximation) and Audio Signals. The attendant slides (with sounds embedded) can be found here. Beware the document is 99 MB large. When I last talked with Bob we did not really have a good sense on how compressive sensing could be applied to audio. In the audio world, sensors of all kinds have been developed and recording large amount of data is not a significant issue as there is a very large pile of money behind it. Besides some of the path shown by Jort Gemmeke for missing data imputation, one wonders in what direction the audio world would allow for a niche to exist and eventually strengthen. The fact that the presentation is 99 MB is a sure sign to me that some improvement can be made. While dictionary learning is a worthy goal, I wonder if compressive sensing would not allow for other type of information to be extracted from signals. For instance, while Blind Signal Separation is an item of intense study, I have not seen anybody use voices signals to infer the physical location where the recording took place in some sort of advanced echolocation technique. Can you imagine if we could see the face of Orson Wells when he was running his War of the Worlds radio show ? Some blinds do it already.



Furthermore, one wonders what is the simple biological basis for a system that allows some moths to jam some bat's radars ?

On a totally different note, the IJCAI-09 papers are located here.

3 comments:

Anonymous said...

I've worked with time-series which had a low sampling rate but recorded a phenomena with components at higher frequencies. This gave a crude image of what was going on. The situation is similar to a sound recording that is not sampling fast enough to pick up high-frequency sounds.

If compressed sensing could be used to retrieve some of the higher frequencies, that would be quite a catch. But I don't know if compressed sensing is any use when, as in my case, the samplings were made at fixed intervals and not randomly?

Igor said...

The work performed on A2I clearly tries to catch some of the high frequency while recording at a low rate. The catch is for them to be able to have a switching mechanism at high frequency so that the low frequency recording is the realization of a CS measurements with the high frequency switching mechanism performing the random modulation. Does that help ?

Now the question is: is there a fast and cheap switch mechanism that can enable this random modulation ?

Igor.

Anonymous said...

Ah I see. I'm out of my league here, but suppose that the low sampling is taken at an interval of say 1 sec plus minus 0.02 sec? It would not be an even random distribution of sample points, but for some parts of the frequency spectrum it might be enough?

It's far from ideal of course, but the crudeness of the devices and variations in the delay between the process and the measurement device _could_ introduce a wider distribution of the true sample interval.

Printfriendly