At meetups and other places, I am always surprised by how the idea of science and technology being at the center of epic fights of ideas is not more widely known. I think it probably has to do with the lack of scientific background most science reporters have. Then again, not everybody is really that interested in hearing about those battles, so it might be just a bias. In the end, most people are customers of technologies and rarely care about how it got in front of them. Yet, it is a story, I should have described in Game of Thrones and the History of Compressive Sensing.
A year ago, I wrote about the steamrollers, (see Predicting the Future: The Steamrollers) i.e a set of "laws" that we are all following without realizing it. The resurgence of neural networks in recent years or simply the ability to do complex Bayesian computations is just an outgrowth of this phenomena. On the hardware side of things, I commented earlier (see Do Not Mess with CMOS) that it is always risky to bet the farm on a technology that competes with one of these steamrollers, i.e. CMOS. And yesterday, I noted Vladimir's feed showing off a video of the IMEC hyperspectral camera.
From Vladimir's text:
This technology is said to enable new possibilities for non-scanning, real-time acquisition of hyperspectral image data-cubes at video-rates.
We've seen a similar set up here on Nuit Blanche, namely a video summary of the work on coded aperture ( Video Compressive Sensing by Larry Carin , compressive hyperspectral camera and Compressive video ). What is the difference between these two approaches ? They both use CMOS technology ! well not exactly in the same manner, see, one is fully using CMOS with no care for signal compression while it expects economies of scale to build very large FPAs, while the other restricts itself to a certain CMOS size and expects an intelligent approach to get more information.
In the IMEC camera, the signal is directly acquired and stored in memory while in the aecond approach, we yearn for the coded aperture to reconstruct the hyperspectral cubes faster than a blink of an eye. In effect, the compressive approach is contingent on having faster reconstruction algorithms or in simpler terms, is the embodiement between dumb economies of scale and an intelligent approach to dealing with data.
In the end, the race for getting faster solvers is all about winning a war. As spectators who have sometimes invested too much of their thought processes in a particular technology, the winning of a dumb CMOS solution might trigger the same reactions as those of numerous watchers of Games of Thrones's Red Wedding (do not click on the previous link nor watch the video below if you haven't seen Episode 9, Season 3)
Which one do you think will win ?
PS: I am not saying that CS cannot be included in the IMEC camera, it surely can but if the architecture is easy to build, there is no reason to expect that or similar cameras get more information through CS, as they will always look to increase the silicon instead of bettering the algorithm.
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