Object classification through a single-pixel detector • Hacker News Board

Re: Object classification through a single-pixel detector

#11

Post by dslkfjg » Thu, Apr 01, 2021, 5:24 PM UTC

The article is confusing. It appears they're leveraging implicit information in the spectrum of the light reflected from objects to settle on the object's identity.

Then they're encoding that information using the light spectrum again - but not the same wavelengths or information as they used to detect them in the first place.

So you have two things being run together. One is the identification of objects and the other is the encoding of information about those objects so that a single pixel detector is all that is needed to read what has been encoded.

Leverging the light coming off objects and processing it using diffusion filters to "learn" or "reason" about what the object is is a sort of side channel attack on the problem of object identification wrt to current approaches.

Encoding thngs into a light so only a single pixel is needed to carry information is an encoding trick.

If I misunderstood anything- entirely possible- feel free to correct.

Re: Object classification through a single-pixel detector

#12

Post by dcanelhas » Thu, Apr 01, 2021, 6:07 PM UTC

jcims wrote: post #6

gmueckl wrote: post #5

Takeaway from roughly skimming the article: the single pixel detector is a plasmonic nanoantenna, which seems to be tunable. This is placed behind a set of optical diffraction gratings that have been optimized to the task using machine learning and essentially transform the spatially encoded information in the input into a spectral encoding. Both the gratings to encode the information and the decoder have been design…

Presumably you could invert it as well and generate images from a single ‘color’. Create a passive spectroscope that converts the observed spectrum to text identiying the chemical compounds it represents.

Alternatively a very complicated way of building a film projector.

Re: Object classification through a single-pixel detector

#13

Post by dcanelhas » Thu, Apr 01, 2021, 7:20 PM UTC

gmueckl wrote: post #5

Takeaway from roughly skimming the article: the single pixel detector is a plasmonic nanoantenna, which seems to be tunable. This is placed behind a set of optical diffraction gratings that have been optimized to the task using machine learning and essentially transform the spatially encoded information in the input into a spectral encoding. Both the gratings to encode the information and the decoder have been design…

I interpreted the result as being the convnet equivalent of implementing a multilayer perceptron network as an analog circuit (e.g. using operational amplifiers).