Cool work! Something that worries me with PCA is that it's designed to retain variance, but variance might nor be the right metric for the semantics we want to retrieve. Ditto UMAP/tSNE that retains distances in lower dimensionality... If semantics are mostly encoded as directions on subsets of dimensions, PCA and friends would be too blunt of a tool... I wonder if a better approach would be linear probes or other decoders for a wide range of the concepts one wants to retrieve, and then optimize compression while keeping those retrievals as high as possible... i.e. tune the compressor to the usecase, like MP3 or MPEG do.
Armchair comment. I would LOVE to be a grad student again and try to pair it with ultrasound speaker arrays, for medical applications. Essentially a super HIFU (High-Intensity Focused Ultrasound) with live feedback. https://en.wikipedia.org/wiki/Focused_ultrasound
howdy! Clay makers here. Can you share more? Did you try Clay v1 or v0.2
What image size embeddings from what instrument?
We did try to relate OSM tags to Clay embeddings, but it didn't scale well. We did not give up, but we are re-considering ( https://github.com/Clay-foundation/earth-text ). I think SatClip plus OSM is a better approach. or LLM embeddings mapped to Clay embeddings...
You'll need to subscribe to Alexa weather plus, for only 9.99$/month.
Now seriously, yes, hyperlocal short-term weather forecast should be a commodity, even public utility?
Can we help you?
We build the equivalent for land, as a non-profit. It's basically a geo Transformer MAE model (plus DINO, plus matrioska, plus ...), but largest and most trained (35 trillion pixels roughly). Most importantly fully open source and open license. I'd love to help you replace land masks with land embeddings, they should significantly help downscale the local effects (e.g. forest versus city) that afaik most weather forecast simplify with static land cover classes at most. https://github.com/Clay-foundation/model
The Planetary Computer from Microsoft has an explore tool, which includes Sentinel-1 at any time and location, rendered on-the-fly. Free open access. No account required.
I fondly remember how a small DC startup subsubsubcontracted to do the front landing page was able to ensure that at least people landed on healthcare.gov page and not a 500 error page.
All thanks for a single server (with one backup) using Jekyll, if I remember the story correctly.
Down the river, right after DC its the Naval Research Laboratory, in front of Alexandria. I worked there, and some crazy folks from Alexandria sometimes crossed the river in canoes to the NRL pier... That got terminated after 911, so they had to bike or drive all the way up around DC and down through Anacostia Air Force Base. Quite a much longer commute.
You can use this link to see the appalling scope of this flood.
The layers you can toggle in the bottom right: 1) Lastest radar, 2) Last year radar for reference, 3) nighlights as a proxy of population.
I selected radar because it's really good at detecting standing water on the ground (as blue).
The part that is confusing is highest resolution (1) full-disk and (2) outer atmosphere: (1) "Full-disk" is clear to understand: the higher the resolution, ^2 the work to make it also full-disk (especially when the Sun rotates differentially and evolves in high-cadence, so you gotta be fast. (2) "Outer atmosphere" is also tricky as only few wavelengths see the outer atmosphere. The vast majority of the light comes from the "surface" or photosphere (hence the name). In this case surface, the highest resolution is roughly 0.05 arcsec or 50km/pixel. But to see the outer parts, you have to do to emission of elements like Iron that only emit when highly ionized and super high temperatures (those are the special characteristics of the sun's outer atmosphere... yes, it's way hotter than the surface, just WAY less dense). Those emissions happen in the Ultraviolet, 17 nanometers, like the caption says. That's like 50 times smaller wavelength. Angular resolution is proportional to wavelength (1.22*wavelength/Diameter) which is on the order of 1000 km/pixel (but linear resolution makes less sense since the atmosphere is such a 3D shape... it's better to say 1 arcsec of resolution).
I might be too biased (I'm a solar physicist) but the explanation above makes the image way cooler and they should have added it): The most detailed image of the Sun's metal corona :D
Microsoft "AI for Earth" | 3 roles in GIS+ML+ Sustainability | ONSITE,REMOTE,VISA all ok
We are building our commitment of the "Planetary Computer" [1]. We are looking for a principal architect (most senior position), a Datasets/ETL senior engineer, and an applications engineer[2]. Candidates for all three roles in the intersection of GIS/Cloud/OSS/ML/sustainability. Asymmetric candidates on these skills ok.
I'm surprised there was no mention (yet!) to fast.ai here. I've decided to learn deep learning this year, after many failed tries with other approaches. Their framework (built over PyTorch), their course, and their community around it are simply the best I've found so far. Very much recommend to anyone who knows a bit of coding and wants to learn Deep Learning quickly and pragmatically.
If you are looking for an alternative, I'd recommend checking out Eversend. (Disclosure, my GF is a co-founder). Much smaller company, but trying to grow it right (malmo FastTrack, Berlin tech starts, ...
OpenCollective [1] is a possible solution. It provides a way to support (monthly contributions or dedicated campaigns), but also the whole back-end of tracking the money and spending, and at least some degree of legal support with paperwork, invoices, ...
I'm surprised there's no mention to adaptive optics in solar physics. It's essentially the same, but some interesting differences. Since there's more light you can have more corrections per second, better approaching the assumption of constant deformación during between corrections on a small view angle ("constant isoplanatic patch"). Also there's no "perfect star" to correct to, so calculation run on a closed loop to basically improve contrast of the reference. The needed correction is also bigger during the day, since the atmosphere is more turbulent. In the last few years there also have been really cool improvements to account for different layers of the atmosphere and better faster algorithms to correct wider and wider fields of view.
This is an example (from my PhD) of the state of the art 9 years ago, but illustrates the huge difference: https://youtu.be/x3JkjXco6m0
Why not just weighting the value of each like by the number of likes given over a period of time? People who spare the likes will give more value, those who like it all, won´t really add much value with their like.