(Disclaimer: I worked in retinal imagery AI for a few years)
I understand why this seems like pseudoscience, but I'd like to explain the assumption better.
As the article states,
> a scan of the retina is the only non-intrusive way to view layers of cells below the skin’s surface.
As an extension, you could also make the case for being able to view the nervous system via such imagery. As stated in [1],
> Despite its peripheral location, the retina or neural portion of the eye, is actually part of the central nervous system.
The theory is that such a non-invasive of visualizing crucial parts of the inner body opens up a "window" into noticing such biomarkers well before the common-but-late-stage symptoms are noticeable.
As for your gripe about:
> let’s dump medical records and retina images into a neural network and see what correlates
, I agree that this requires more careful analysis, but I assume that will happen as part of clinical trials of any such technology. The first foray, which is more experimental than anything concrete, is what this article seems to suggest.
Generative AI for images and music produce pixels and waveform data, respectively. I wonder if there is research into "procedural" data; so in this case, it would be SVG elements and, perhaps, MIDI data respectively.
I know training data would be much more harder to get, (notwithstanding legal ramifications), but I think that creating structured, procedural data will be much more interesting than just the final, "raw" output!
They raised $580M in this Series B round, which is interesting:
They provide links to 5 papers on their website [0], but they are either ArXiv links (so preprints, not peer-reviewed) or blog posts on a website that seems to only contain their own material. They do not cite any clients on their website, they do not showcase any product that they sell.
For a company with 40 employees, only over a year in existence, and with little to no "verification" of their work, I wonder how do they raise such a huge amount?
I think that, as of this moment, the tool does not take the user's current region into account. I know two songs that are very similar[0][1], but from vastly different regions, and I could easily get them both detected by this tool just by humming slightly differently.
I am a non-native English speaker, and I have been trying songs from different countries. I did notice that English songs have a better match than songs from other regions and languages[0]. I wonder if their training dataset has "overfitted" to such music, or is such music inherently represented by some underlying features that are better distinguished than others.
[0]: E.g. English ("My Heart will Go On" and "Skyfall") fit with 78% and 85% respectively, while Japanese ("Tonari no Totoro") and Hindi ("Tum Hi Ho") fit with 42% and 48% respectively
> ... in the era of weapons-grade language models?
Just an aside: Although common perception is that "weapons-grade" refers to the highest quality, in reality it refers to the lowest-possible quality that meets the bare minimum requirements!
I came across [1] some weeks ago, which (although didn't give me the flexibility to choose my topic), did fulfill my need to look at the view from both sides on, presumably, the hottest topic of the day.
They have a "curation"-method to selecting the articles to show. I am curious as to if this extension uses pre-selected websites (e.g. WSJ will always show the Right-leaning view), or is there something smarter than that
I think the issue people have with these names is that their English meanings (as you described) make sense when the positive class is not as prevalent as the negative class. If they are equally probable (or worse, if it's the opposite), then the English meanings quickly become out-of-context
Not much on there, but hey, something's better than nothing!