Friday, December 29, 2006

Nothing short of a revolution. Part I: When a scam can kill you

This story is about the revolution brought about by compressed sensing or compressive sampling. It is a beautiful story as far as I can tell and I will try to parse it in three or four parts. I write it as an outsider and I therefore hope to be forgiven if I missing names.

It all started with the (re)discovery of wavelets by Jean Morlet and the work of people like Alex Grossman. The field was expanded by people like Yves Meyer, Stephane Mallat, Ingrid Daubechies and many others. As pointed out in the text about Jean Morlet, it all started badly:

...Until now, his only reward for years of perseverance and creativity in producing this extraordinary tool was an early retirement from Elf...

Elf was not fearing knowledge or fearing to be more competitive but rather the company had been hit with a scandal that nearly shattered a government and a presidency. The avions renifleurs affair (sniffer planes) made most of the management at ELF very cautious/squeamish about any type of extraordinary scientific announcement. So when Morlet came in and told people about wavelets, they had him retire. While the story is cute, it took a good ten years before wavelets became of interest in the generic engineering literature.

As an engineer myself, the turning point for making these functions useful was when Dave Donoho started to make available a series of routines on the internet so that other researchers could reproduce his own published material (David Donoho wants his research to be reproducible because he thinks that publication is not scholarship but merely an advertisement for it). The Wavelab toolbox was revolutionary because, in effect, since most of these wavelet functions were not analytic (they are mostly defined as a solution to a recurence relation) most engineers had to first build something to get to see a "wavelet" function. And while there were wavelets built from spline functions by Chui, Goswami and Chan these never really took off in the literature. Wavelets are now routinely used in every areas of engineering and have joined Fourier analysis in the series of tools engineers and scientists use in their daily lives. They have changed the face of applied harmonic analysis and change our ways we look at signals such as image, voice, etc. However, the revolution only starts here. The bigger boom came from another area of mathematics...

Thursday, December 21, 2006

Reflecting on a moving view

I recently flew on a small shuttle plane between College Station and Houston and wanted to see what the view looked like in a panorama. Stitching several photos together using Autopano Pro, this is what I eventually saw from 4,000 feet. This is a little bit different than say 112,000 feet since the angle to the scence of interest makes it difficult to patch images together. When the camera is close to the land, the movement of the plane changes quite rapidily the angle to the scene and it then becomes difficult for the software to patch images together.

Wednesday, December 20, 2006

Thought compression - part deux: A python perspective

Since I last wrote about thought compression, the Hutter prize was launched and somebody has already found a way to earn part of that prize. The concept is to find a compression algorithm that compresses a 100 MB from Wikipedia in a lossless fashion to less than 17 MB. 17 MB is the benchmark value for the best "dumb" compressors like zip...The idea is that using relationships between words, one is very likely to be able to compress further a text as opposed to relying on a smart but really dumb way of compressing words of a dictionary based on the frequency use. In the end, the sense is that artificial intelligence can benefit from this because it would permit the "comparison" between statements using distances such as the ones devised by Rudi Cilibrasi.

While compressing text is fine, another area of compression still impresses me. The use of a very expressive programming language that nearly reads like pseudo-code. Python continues on providing examples as to why it falls in this category of expressive languages. For instance, the p2p example was nice but others are similarly intriguing like:

- the pageRank algorithm

- an IRC bot

- a Markov Chain Monte Carlo solver (MCMC solver) or

- Autonomous car programming in the past and in the present.

The MCMC code one is more intriguing than anything else. Because of their ability to compute multidimensional integrals, MCMC methods have allowed the use of Bayesian statistics and Bayesian programming in problems of inferences, a hallmark of artificial intelligence that the Hutter prize is seeking to enable.

Sunday, December 17, 2006

Diabetes and Central Nervous System breakthrough

A recent study shows that Diabetes seems to be linked to the nervous system. This is indeed a surprise. However, the most surprising fact is that it was not discovered earlier as it is known that mental illness is linked to diabetes. More specifically, according to several studies dating back fifty years ago we could see the following :
... based on a retrospective review of medical data for 569 randomly selected patients with the two disorders admitted to a state psychiatric hospital between 1940 and 1950—before antipsychotic medications were available-found that metabolic disturbances were significantly greater in those patients than among the general population...

Friday, November 24, 2006

We are slowly getting there

For the past two months, we have been talking to one of the maker of Autopano Pro so that we could process a very large panorama of about 609 images (each of which is 3.2 Mpixels large). The current beta release (version 1.3 RC 3) still has a memory leakage problem for very large panoramas like this one. This monster is about 2 GB large. One of the smaller panorama can be found here. Others can be found here.

Friday, November 10, 2006

Google can save your life

In a recent study, it was shown that good professionals could use Google in problems with hard to figure diagnostics. This is hardly a big discovery to the average person. On the other hand, I know first hand, of somebody who used Google and figured from the first three hits that he was having a big problem that could only be readily remedied by not staying in a high altitude environment and aiming straight for a hospital for treatment.

A C compiler for GPUs

While GPU programming might be fun on its own, the 4 to 5 times increase in speed is not likely to draw a whole slew of people into it. One of the reason may be that if you have to learn a different language, you might as well go straight for an FPGA instead. NVIDIA seems to have figured this out and they just came up with a bridge between the general community and GPU by announcing CUDA a C compiler for GPUs.

Thursday, November 09, 2006

Space Vantage Points

When Nadar went up his balloon 150 years ago and took the first aerial photo, he triggered Jules Vernes into writing "Five weeks in a balloon".
The first panoramic view of a natural disaster was taken one hundred years ago after the 1906 San Francisco Earthquake using a kite.

The first photo taken from space was obtained in 1946, from a V2 rocket and the first panorama in space was taken two years later.

Fifty years later, thanks to the HASP platform, we took one of the longest panoramic view from a high altitude balloon. This is mainly due to our ability to store large amount of data in common cameras (4GB).
. The interesting aspect of our approach relies on the fact that our low cost camera does not need to be equipped with either a GPS or an Inertial Navigational Unit: The use of a software like Autopano Pro enables us to patch automatically all of our pictures together and create a single large map.

Friday, October 27, 2006

Eye tracking and Autism: part IV

Here is a cheaper implementation of an Eye tracker. The details can be found on a wiki set up by Derrick Parkhurst. We are implementing one of these but have had some problems with cutting the chips out of the board and are about to start some test on it.

Monday, October 02, 2006

The slashdot effect

This Slashdot announcement for the preliminary results on GeoCam generated about 25,000 viewers in the course of two days. While the comment section on Slashdot is not that useful, we received some very helpful tips and ideas on our blog. I have stitched together some of these panaromas with one of the non-free version of the stitching algorithm and I am very impressed. The current panoramas are 25 to 30 MB large.

Tuesday, September 26, 2006

Cognitive Convergence

I am not blogging that much these days for several reasons. First, I am looking into how we can use some of the artificial intelligence techniques we are developing for the DARPA Urban Grand Challenge to diagnose and explore autism. I recently attended a talk by Hideki Kozima who uses of small robots like Keepon to evaluate the socialization of autistic kids. Take a look at the video here. After five to ten sessions with the Keepon, he could show the beginning of a joint attention development in autistic kids. This was very impressive.

I am also involved in the processing of the data we just received from our GeoCam on HASP. We have 4 GB of data to share with the rest of the world. We are developing a strategy on how to do this efficiently.

We are considering a run for the DARPA Urban Grand Challenge with Pegasus Bridge 1. Only this time, the mechanical aspect of the project will take a back seat to other technologies we are developing. We are interested in driver's gaze recording and supervised learning of road driving behavior. But more on this later...

Monday, August 28, 2006

Close enough

The issue of designing a collaborative task manager needs to also take into account real data from previous experiments where collaborative behavior took place. One of the issue is the sizing of the application. Soon into our experiment (in 2002), we jokingly came up with the DC law that stated that the size of the conversation threads between collaborating parties was 1/1000 th the size of the documents attached to these conversation threads. As of today, four years later, the size of the "library" of documents is 22.17 GB. The size of the conversation threads is 18.66 MB. Close enough. This is somewhat amazing that the "law" still holds. Indeed, in a matter of four years, new formats have appeared in the archives: MPEGs or very large JPEGs, yet the scaling still holds.

Wednesday, July 26, 2006

GeoCam: An Off-The-Shelf Imager for High Altitude Balloons

The blog for the GeoCam Imager is up and can be found here. Here is the most recent photo of it in its casing.

The whole reason as to why we think this off-the-shelf camera is a good idea is explained here.

Fabricating elements around current high end consumer digital cameras (like the clicker) is the most effective way of utilizing advances in digital camera development. We intend on showing how we did all this on sites like Hackaday. Thank you to the HASP program at LSU for giving us a seat in the High Altitude Student Platform (HASP).

Tuesday, July 04, 2006

Designing a Collaborative Task Manager

I mentionned before the fact that multitaskers could not be more efficient than single minded people because there was a cognitive cost associated with having to upload their memory of the previous tasks. In a remarkable study, Gloria Mark goes farther and notes:

What we found is that the average amount of time that people spent on any single event before being interrupted or before switching was about three minutes. Actually, three minutes and five seconds, on average. That does not include formal meetings, because we figured if they were in a formal meeting, they were prisoners at the meeting, right? They couldn't leave or switch activities. So we didn't count that. Then we looked at [use of] devices, working on a PC, the desk phone, using any kind of paper document, using a cell phone. We found the average amount of time that people spent working on a device before switching was 2 minutes and 11 seconds.

But we called it a "working sphere" because "project" has a limited connotation, and a working sphere is a broader idea. Anything where there's a common goal, there's a certain group of people involved with it, there are certain resources attached to it, it has its own time framework and its own deadline, is a working sphere

So even when we took out what we call "non-significant" interruptions, we find that people still worked 12 minutes and 18 seconds in a working sphere before switching.

But there are also internal interruptions; for whatever reason, people interrupt themselves of their own volition and switch to something else. And what fascinates me is that people interrupted themselves almost as much as they were interrupted by external sources. They interrupted themselves about 44% of the time. The rest of the interruptions were from external sources.

GMJ: How long does it take to get back to work after an interruption?

Mark: There's good news and bad news. To have a uniform comparison, we looked at all work that was interrupted and resumed on the same day. The good news is that most interrupted work was resumed on the same day -- 81.9 percent -- and it was resumed, on average, in 23 minutes and 15 seconds, which I guess is not so long.

But the bad news is, when you're interrupted, you don't immediately go back to the task you were doing before you were interrupted. There are about two intervening tasks before you go back to your original task, so it takes more effort to reorient back to the original task. Also, interruptions change the physical environment. For example, someone has asked you for information and you have opened new windows on your desktop, or people have given you papers that are now arranged on your desk. So often the physical layout of your environment has changed, and it's harder to reconstruct where you were. So there's a cognitive cost to an interruption.

The designers of a task manager for a small business could learn a lot from these results. For instance, in my case, the application we developed at my workplace was aimed at handling the communication load of about 60 people with many different schedules, projects and from different functional groups. After some time, we noticed certain behaviors. First, we found out that if the application was not used by everybody uniformly, sooner rather than later, an assymetry of the information stream helped people who were least likely to contribute/use the system. In effect, cooperation between users was less effective over time. The second finding was a little more subtle. The application was written in PHP but did not use AJAX. Every time the page would load with new information, it took the new page about 10 seconds to refresh. In light of the study by Mark, it seems that this is enough of a burden that this cost of "waiting" for the refresh would have people drift into other applications and other "working spheres". For participants, the time spent in updating knowledge in the application felt like "feeding the beast" as opposed to being part of a normal cognitive process.

Wednesday, June 21, 2006

Tex-MEMS VIII (or Tex-MEMS 2006) will take place in Dallas

It looks like a fire that occured in San Antonio is precluding the organizers to have tex-mems VIII in San Antonio this year. The folks at UT Dallas have decided to organize it instead. The web site will be up shortly.

Wednesday, March 29, 2006

Going high

Our proposal for the Geocam has been selected to fly on-board a high altitude balloon campaign organized by the Louisiana Space Consortium. It will be launched from Fort Sumner, NM. Movies taken from high altitude balloons can be found at the University of Montana within the BOREALIS High Altitude Balloon program.

We expect to use an JVC Everio Hard Disk Camcorder in order to fit our tiny requirements.

Friday, March 17, 2006

Seeing through takes time

When Mikhail Vorontsov talks (Anisoplanatic imaging through turbulent media: image recovery by local information fusion from a set of short-exposure images, [J. Opt. Soc. Am. A/ Vol. 18, No. 6/June 2001]) about superresolution, he is basically saying that the turbulence located between the camera and the object of interest will help in focusing. The idea is that turbulence will bend rays of lights and allow for micro-lensing to occur. It is interesting that in effect, this idea cannot be applied to Space for short distances. In Space however, while there is no turbulence, one can rely on gravity to bend light and allow for this micro-lensing to occur. This is how the latest Earth like object was found. The technique itself has to figure out which frames and what part of the frame is the one of highest quality. This is no small feat. In effect, one has to determine a criterion that evaluates the quality of focus for every part of the frame and has to perform the fusion of these "lucky" parts of the frame.

Saturday, February 18, 2006

Multi-use of Google Maps

Google Maps Mania just featured Peter Pesti at Georgia Tech with his Google Maps Nighttime!. At the same time, Google also uses Google Maps to geographically target customers through their

Adwords program (here some of the screen grab of the application within Adwords).

On another note, if one were to have a very high resolution camera, one could shoot a high definition picture and use the same Google Maps interface to go into the details of that picture (like in the following game using Google Maps). More information on how to introduce a custom map into Google maps can be found in GoogleMapki. This possibility offers other avenues,. For instance, if one were to have an HDTV camera, one could patch different pictures from a scene (every picture is taken at 1/30 th second) and have the Google interface allow navigation through it. One cannot but think that a similar system would work well for Nat's world (a program designed to help Autistic kids know their environment by navigating through it.)

Thursday, February 02, 2006

Safety is our job Number...errr... make that Number 6

Some time ago, two tankers collided off the French coast on the English Channel. Here is what happens when corporate speak does not mean what it should mean. This tanker sank with all its phosphoric acid payload. (here is another example)

Tuesday, January 31, 2006

Optical Illusion

Here is a very interesting optical illusion borne out of the lack of other types of cues than the one provided by the planes themselves yielding a lack of depth perception. Depth perception is not just coming from stereo (two eyes) imaging, it is a combination of texture, colors that allows the brain to evaluate depth. In this case, all these information are lacking yielding this impression that the planes are about to collide.

Saturday, January 28, 2006

Friday, January 27, 2006

Imagine: Autonomous Blimps

I just saw this on Lemonodor and I thought it was cool: Autonomous blimps. They just need to put a micro-camera on these.

How long before we have a weight loss program based on this ?

Researchers at UT have found that if you stay in darkness about 48 hours, you begin to burn fat.

Thinking outside of our world

Adrian Bejan, a noted heat transfer researcher, believes that he has found laws that would explain locomotion on the ground, in the air and in water. This is pretty significant since it does not seem obvious on how to connect swimming behavior to walking on the ground. Most of his argument revolves around what he calls constructal theory which as far as I understand says that nature allows for different shapes to occur as a result of optimizing heat transfer at different scales. That results allows for a clear explanation of tree like structures for the lungs, veins, trees, roots which has mostly being looked at as being an ad-hoc assumption. This is a also a result that explains why Nature is not made of fractals. In 2000, he showed you could draw a line through meat flies on 747s on the same graph (the x-axis is the mass, the y-axis is their theoretical speed). This time he shows how, given gravity, density of the body, air and water he can fit pretty much all living things in log-log straight lines.

This is very interesting on many levels. Obviously, our brain says it OK when we see shapes that follow these principles:

  • In the MEMS world, gravity becomes less important compared to Van der Waals forces. Why should normal shapes found at the human scale world be found at these microscales ?

  • On other planets, such as Europa, where there is an expectation of prospects of life, how does different gravity levels (1/6th of a g) changes the shapes of the living bodies there ? Similarly, while most current spacecrafts look like airplanes or cylinders (rocket parts), is there an optimum shaping mechanism in zero-g, how will future spacecrafts built in space look like ?

  • Can we make out the muscle structure of dinosaurs from these types of consideration ? and what is the largest animal we could ever find when digging since every year, there is an announcement of finding larger and larger prehistoric animals ?
  • Return to the Future

    From Hackaday, Bogdan Marinescu shows off his inventive ReVaLuaTe which stands for Renesas Valuable Lua Terminal. This a LUA based microcontroller. The catch: no PC required.
    It uses the the high level LUA programming language. This looks like using BASIC on the Sinclair ZX Spectrum computer.

    Tuesday, January 24, 2006

    Tex-MEMS VIII in San Antonio

    Tex-MEMS VIII will occur in San Antonio on September 21-22, 2006. The conference will include presentations of some of the most prominent experts in the fields of Microfluidics, Microfabrication, Sensors and Nanotechnology. Check their site out.

    For information, previous meetings were held at the following locations/dates:

  • Tex-MEMS I : August 23, 1999, Texas A&M University, College station, TX

  • Tex-MEMS II : May 6, 2000, Southern Methodist University, Dallas, TX

  • Tex-MEMS III : Jun 7, 2001, UT-Dallas, Dallas, TX

  • Tex-MEMS IV : July 11, 2002, Texas Tech University, Lubbock, TX

  • Tex-MEMS V : May 6, 2003, UT-Arlington, Fort Worth, TX

  • Tex-MEMS VI : September 9, 2004, Texas A&M University, College Station, TX

  • Tex-MEMS VII : September 21-22, 2005, UT-El Paso, El Paso, TX

  • Tex-MEMS VIII : International Conference on Micro & Nano Systems, Spetember 21, 2006, UT-San Antonio, San Antonio, TX
  • Saturday, January 14, 2006

    Eye tracking and Autism: part III

    Hackaday pointed to an effort on how to build an eye tracker by Jeff Pelz at RIT (paper here.) This is problaby a good way for obtaining a larger amount of data on autism and eye tracking than currently exist. I have had entries on this before (part 1, part 2).

    Friday, January 13, 2006

    Bayesian cognition: a core subject

    This coming monday, a workshop will take place in Paris on Bayesian Cognition. There is an interesting list of speakers. To name of few: Sebastian Thrun, the winner of DARPA's Grand Challenge, Rajesh Rao, Pierre Bessiere (Cycab project), Daniel Wolpert and others that I am sure I will mention in this blog...