Friday, June 07, 2013

This week in Review: COxSwAIN, Machine Learning and Sensors MeetUps, SigFox, Lensless Single Pixel camera, Around the Blogs in 78 hours

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:

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: 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:

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.

This past week, we've also seen a number of interesting technical blog entries:




While on Nuit Blanche, we covered:

No comments: