It's good to see Google in the last couple of weeks launching a bunch of [1]projects that are more in line with their mission of 'organising the world's information'.
From the video, it appears this is an AI-plotted hugely multidimensional space t-sne'd onto two dimensions.
Would be interesting to do some kind of ML on how best to present hugely multidimensional spaces onto two _interactive_ dimensions. Where one AI is deciding how things are projected and how it can be manipulated, and another AI is limited to some virtual "mouse, keyboard, 2D screen" to make inferences. Such that it's optimized for faster, more correct inferences.
Unfortunately they only hint at the envisioned application in the video and don't provide any further links, but the idea is amazing: Use sounds to monitor bio-diversity. Imagine we'd not need cameras and lots of luck to "catch" proof of an animals existence but a grid of interconnected omnidirectional microphones. We'd get real time tracking of individual animals in 3D and could have smartphones literally point the way to yet uncatalogued or even undiscovered species.
Somewhat related, but there was a very good talk at MLConf 2017 about using sound to catch illegal logging in the Amazon. Similar premise: collect the sound, analyze the patterns, and classify.
I worked on this as part of a project with US Fish and Wildlife in 2015 for bats! We were monitoring local populations to track the spread of White Nose syndrome. It wasn't in 3D, but we linked the audio data / classifications with lat/long points and plotted population estimates on maps.
As a birder, this looks like a failed experiment to me. Or I don't understand what their goal was. The groupings make little sense in terms of what these species sound like. I'm guessing that's an artifact of the way they sampled the sounds, losing macro properties. Kind of like grouping the words 'paramour', 'enmity' and 'hamster' together bc they all contain /m/ sound.
There are some instances of the same bird in multiple locations (great horned owl). Presumably multiple recordings of the same bird. My initial reaction to them not being neighboring is to wonder about the quality of the result. Maybe better feature engineering needed to make this biologically relevant. Any other interpretations?
It's interesting, at the very local level I think most humans wouldn't think two adjacent bird sounds are all that similar. But if you drag along a long line and listen to a series of different birds you can "hear" a definite organized progression that seems to be organizing rhythm and major tones into groups.
On touch screen it is very BBC Radiophonic Workshop, a nice toy to get some sounds from, to suit a play on Radio 4, that sort of thing. So you could make lots of 'instruments' by applying AI to organising sounds into some space that can be used as per this example and a touchscreen. It my make ornithologists cringe but I think the merit in this work could be elsewhere, as a creative musical instrument of sorts.
It would be interesting to see how similarity in bird call behavior tracks (or doesn't) the phylogenetic relationship between species. My hypothesis would be that bird calls are influenced by other birds in the same ecosystem (imitation or differentiation, and reflecting a high degree of cultural learning) rather than the null hypothesis of genetic transmission.
43 comments
[ 3.6 ms ] story [ 85.0 ms ] threadThe tiny images are just spectograms/fft as far as I can tell.
edit: it's very fun though to click+drag, haha
[1] http://www.smithsonianmag.com/smart-news/google-digitizes-30...
Would be interesting to do some kind of ML on how best to present hugely multidimensional spaces onto two _interactive_ dimensions. Where one AI is deciding how things are projected and how it can be manipulated, and another AI is limited to some virtual "mouse, keyboard, 2D screen" to make inferences. Such that it's optimized for faster, more correct inferences.
> Oops, sorry for the tech trouble. For the best experience, view in Chrome browser.
But then, I block some stuff: ads, WebGL, WebRTC, and tracking.
It has the most unique call of them all.
Does anyone know what they are doing t-SNE on? i.e. are they just doing t-SNE on the raw waveforms? Or the MFCC spectrogram? Or what?