Thread by @EricTopol

Most deep learning #AI work to date has been supervised. 1. Now we're seeing a trend towards much more unsupervised, aka "self-supervised" (after @ylecun) or weakly supervised.
Figure from @GaryChoy in #DeepMedicine

@ylecun @GaryChoy 2. In the past 2 weeks, 4 important papers published:
—Process 3.3 M abstracts in materials science
—Prevail at 6-player Texas Hold'em poker
—Solve Rubik's cube in 1.2 secs
—Classify whole slide #cancer pathology images

@ylecun @GaryChoy Links to these reports:
1. nature.com/articles/s4158… @nature
2.science.sciencemag.org/content/early/… @sciencemagazine
3. nature.com/articles/s4225… @NatMachIntell
4. nature.com/articles/s4159… @NatureMedicine

@ylecun @GaryChoy @nature @sciencemagazine @NatMachIntell @NatureMedicine 4. And this week's one on heart MRI images to classify 4000 aortic valves (bicuspid or tricuspid)
nature.com/articles/s4146… @NatureComms by @jasonafries @JamesRPriestMD
and colleagues @StanfordMed @StanfordEng

@ylecun @GaryChoy @nature @sciencemagazine @NatMachIntell @NatureMedicine @NatureComms @jasonafries @JamesRPriestMD @StanfordMed @StanfordEng 5. Big, labelled datasets in medicine are few and far between. Like this one, MIMIC, used by >12,000 researchers in 70 countries for ~500 publications statnews.com/2019/07/12/bos… @statnews by @RebeccaDRobbins The same, limited datasets are used time and time again.

@ylecun @GaryChoy @nature @sciencemagazine @NatMachIntell @NatureMedicine @NatureComms @jasonafries @JamesRPriestMD @StanfordMed @StanfordEng @statnews @RebeccaDRobbins 6. So, @ylecun is right, but whether unsupervised DNN #AI (what may be considered a "workaround" plan) will be applicable in the near term for #healthcare applications remains uncertain. We clearly are in need a lot more well-annotated datasets (+ multi-modal, not just #EHR).