Cable and I went through a brainstorming session over the past few weeks on how to best respond to NASA's RFP on SmallSat Technology Partnerships FY13 program. The result is a proposal for which Cable is a PI entitled: COxSwAIN COmpressive Sensing for Advanced Imaging and Navigation. What's a Coxswain, you say ? Like you, I had to look it up on Wikipedia: "The coxswain /ˈkɒksən/ is the person in charge of a boat, particularly its navigation and steering". I immediately liked the navigation and pathfinder image it entailed.
In effect, what you notice when you look at the literature of sensors for Cubesats and other picosats is that not much innovation is going on there [1] and one shouldn't be surprised. Those things are limited in power (1 watt or less) and bandwidth and in the minds of most on board sensors are expected to deliver remote sensing products in line with common remote sensing capabilities. Our view in the proposal is that NASA ought to investigate the new trade-offs that come from using compressive sensing and other machine learning/computer vision techniques on those power and bandwidth constrained platforms. We'll see where that goes. [Yes, the image on the left comes from GeoCam.]
In a different direction, I attended two Meet-Ups here on Machine Learning and another one from the Experimental Sensors club. The first meetup had a very diverse crowd and I think it is the start of interesting discussions. The second meetup was focused on Sensors and Privacy. In both meet-ups, I was asked exactly the same question: "What is a sensor ?" and I think I "involuntarily" set myself for a presentation in both groups.
Listening to some descriptions, I feel there is a disconnect between what people call sensors and what sensors really are. Yes, a temperature sensor gives a temperature reading, but what does several temperatures sensors give. With the right architecture, it may provide other information. Genome sequencers are a different example. It is a sensor: Do you have one in your living room, no! Can you get data from one for less than US$5000 absolutely! Do you have a keen understanding of what that means, maybe not. And then I haven't talked about the newer paradigms: With compressive sensing, one can definitely see a convergence between the sensing part and the analysis part (machine learning) and one can wonder at what point, data obtained through these new sensing mechanisms can become much more informative as reconstruction algorithms become more powerful. With pure data, we've seen examples of de-anonymization with the Netflix prize and the rise of advanced matrix factorization such as matrix completion. With sensors, we're now seeing the possibility of performing informative 1-bit compressive sensing and group testing.
For backgrounders, here are some of the recent presentations I made on sensors and compressive sensing:
Listening to some descriptions, I feel there is a disconnect between what people call sensors and what sensors really are. Yes, a temperature sensor gives a temperature reading, but what does several temperatures sensors give. With the right architecture, it may provide other information. Genome sequencers are a different example. It is a sensor: Do you have one in your living room, no! Can you get data from one for less than US$5000 absolutely! Do you have a keen understanding of what that means, maybe not. And then I haven't talked about the newer paradigms: With compressive sensing, one can definitely see a convergence between the sensing part and the analysis part (machine learning) and one can wonder at what point, data obtained through these new sensing mechanisms can become much more informative as reconstruction algorithms become more powerful. With pure data, we've seen examples of de-anonymization with the Netflix prize and the rise of advanced matrix factorization such as matrix completion. With sensors, we're now seeing the possibility of performing informative 1-bit compressive sensing and group testing.
For backgrounders, here are some of the recent presentations I made on sensors and compressive sensing:
- Compressive Sensing: What is it good for ? Supelec, April 24th, 2013 (talk, talk version 2)
- Successes and Failures of Autonomous Things in Space and on Land, Embedded Paris Meetup #2
- Les Cameras Aleatoires ( Random Imagers ) DorkBot, Paris, 2012.
All in all, presentations as a means of exposing ideas are good but I think they are missing the wow factor. Maybe it's just a question of building a simple compressive sensing system to get people's attention. Romain Cochet, who read this recent entry on the Lensless Imaging by Compressive Sensing, and I wondered about the type of LCD screen one would use in order to have a similar mock-up for dog and pony shows. If you have other ideas, let's talk.
During the Machine Learning meetup, I met Gabriel who then reminded me on Twitter about the Johnson-Lindenstrauss Lemma as featured in Scikit Learn. It is here at: http://scikit-learn.org/stable/modules/random_projection.html. Olivier then added that we should check the attendant plots here.
This week, I was also made aware of SigFox, a network allowing communication in sensor networks. The bandwidth is low but the number of towers is low as well compared to CDMA and GSM. All this point to sensors for the Internet Of Things (#iot) with low bandwidth capabilities, a subject of clear relevance to compressive sensing.
During the Machine Learning meetup, I met Gabriel who then reminded me on Twitter about the Johnson-Lindenstrauss Lemma as featured in Scikit Learn. It is here at: http://scikit-learn.org/stable/modules/random_projection.html. Olivier then added that we should check the attendant plots here.
This week, I was also made aware of SigFox, a network allowing communication in sensor networks. The bandwidth is low but the number of towers is low as well compared to CDMA and GSM. All this point to sensors for the Internet Of Things (#iot) with low bandwidth capabilities, a subject of clear relevance to compressive sensing.
This past week, we also saw some frenzy over the Lensless Imaging arxiv paper. The MIT Technology Review ArXiv blog rightfully focused on the lens aspect of the set-up in Bell Labs Invents Lensless Camera. Although one could argue that the lens is the LCD, one could see some information loss when it got picked up by larger outfits:
- PopSci: This Lensless Camera Is Never Out Of Focus , actually it is. What is described here is a time modulated coded aperture. As one knows, the image formed in the black box, is absolutely out of focus. Only the deconvolution produced by compressive sensing can render it out of focus.
- Wired: Augmented Reality: Lensless “compressive sensing” camera
- Time: Finally, a Camera Without a Lens (and a Sensor the Size of a Pixel), actually there have been a number of single pixel cameras (maybe I ought to write a blog entry on this). But the reason it is cringing is that a start-up company like InView is already commercializing single pixel cameras. InView was recently mentioned in the list of start-ups last Friday.
I have tried to provide some insights into the matter on some comment section of these articles. Overall, it is useful for most to understand what compressive does in the context of single pixel cameras. Some of the most interesting insights however are in the comments section of the dpreview piece. For experts, the issue is about extended objects as pointed out by Roummel Marcia, Rebecca Willett, and Zachary Harmany and as featured in Comparing a Single Pixel Camera, a Traditional Coded Aperture and a Compressive Coded Aperture Image of Saturn. The jury is still out though if you ask me.
In hindsight, the extract from Muthu's 2010 Massive Data Streams Research: Where to Go is becoming all the more relevant:
Compressed Functional Sensing. Streaming and compressed sensing brought two groups of researchers (CS and signal processing) together on common problems of what is the minimal amount of data to be sensed or captured or stored, so data sources can be reconstructed, at least approximately. This has been a productive development for research with fundamental insights into geometry of high dimensional spaces as well as the Uncertainty Principle. In addition, Engineering and Industry has been impacted significantly with analog to information paradigm. This is however just the beginning. We need to extend compressed sensing to functional sensing, where we sense only what is appropriate to compute different function (rather than simply reconstructing) and furthermore, extend the theory to massively distributed and continual framework to be truly useful for new massive data applications above.
Dusty
Dirk:
- SSVM day 2
- SSVM first day
- Note to self: Some papers which should not escape from my focus
- Geometry, Imaging and Computing
Suresh
Franck
- Geometric Science of Information (GSI): programme is out!
- Random uniform distribution inside the probability simplex
Danny
- Notable presentations at Technion TCE conference 2013: RevMiner & Boom
- Nice Scala Tutorial
- Cascading and Scalding
Larry
Tim
- Elsevier journals: has anything changed?
- Answers, results of polls, and a brief description of the program
Mark:
- Meet the Bregman Divergences
- Bayesian Updating as Regularised Optimisation
- Machine Learning Research Jobs at NICTA
Michael
Greg:
Vladimir
- Intel Capital Creates $100M Perceptual Computing Fund
- Gesture Recognition Middleware News
- Singapore University Presents Graphene Image Sensor (you really want to read the comments)
Roger:
John:
John:
Ben
While on Nuit Blanche, we covered:
Join the CompressiveSensing subreddit or the Google+ Community and post there !
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.
While on Nuit Blanche, we covered:
- Post-Doc at QM-MPI Joint Information System Lab in Macau, China
- Structure Discovery in Nonparametric Regression through Compositional Kernel Search - implementation -
- Lensless Imaging by Compressive Sensing
- Nuit Blanche in Review (May 2013)
- Engineering a Healthcare System to Deliver Genomic Medicine
- Saturday Morning Videos
- A PostDoc and a Ph.D in Medical Imaging, Machine Learning and Compressive Sensing at INRIA, Rennes, France
- Start-ups: GraphLab, Wise.io, InView, Centice, Aqueti
- SPARS13 Abstracts, ROKS 2013 List of Papers, SAHD2013 and other CS/MF related meetings
- Hierarchical Tucker Tensor Optimization - Applications to Tensor Completion
- qGeomMC: A Quotient Geometric approach to low-rank Matrix Completion - implementation
- One million here, one million there, soon enough we're talking about real readership!
- Sunday Morning Insight: The 200 Year Gap.
- Saturday Morning Videos: ProLand, SIGGRAPH, M7 solar flare, Xbox One and the new Kinect
- Nuit Blanche Reader's Reviews: Traveling Salesman, Quantum Imaging and more, Google +1s, Discussions on Nuit Blanche, G+, Reddit and the LinkedIn groups
[1] A survey and assessment of the capabilities of Cubesats for Earth observation Daniel Selva, David Krejci
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.
No comments:
Post a Comment