Tom Beer

ML for accessible healthcare @KHealth
~ Optimism in the face of uncertainty

[Tom Beer @aTomBeer]

It took a decade of work but here we are
Proud of my colleagues at @khealth

[Eric Topol @EricTopol]

When physicians with access to A.I. were compared with A.I. for virtual urgent care visits
The A.I. performed better.
@AnnalsofIM
https://t.co/PAh75ckpHd

[Tom Beer @aTomBeer]

LLM vibes

[Tom Beer @aTomBeer]

Sparks of AGI

[Tom Beer @aTomBeer]

Exactly what the community needs. Fantastic!

[DagsHub @TheRealDagsHub]

🚀🚨 Launching Cookiecutter-MLOps 🚀🚨 – An open-source project scaffolding for machine learning and #MLOps, built on top of Cookiecutter Data Science, and inspired by @sh_reya’s amazing tweetstorm.

https://t.co/ddCGkgh6IE

[DagsHub @TheRealDagsHub]

Wouldn’t it be cool if interactive #Jupyter notebook outputs would stay interactive after you committed them to git?

[Tom Beer @aTomBeer]

So many ML models for the ICU, so little utility…why?
This work is the first to investigate what are the clinicians’ requirements from ICU prediction models.

Nice work @BarEini and team!

[Bar Eini-Porat @BarEini]

I am happy to share that our paper “Tell me something interesting: Clinical utility of machine learning prediction models in the ICU” was accepted to @JBI_Journal! @ShalitUri @DannyEytan @ofraam

https://t.co/VysohdQw5W

[Tom Beer @aTomBeer]

For me, what’s missing the most is a rule of thumb for determining if some case of overfitting (train error < test error) is benign, or should I select a different model. If you have such a criterion, I would like to hear!

[Tom Beer @aTomBeer]

The bias variance tradeoff - a universal truth or an irrelevant lesson from ML101?

A thread revisiting the classical principle in modern times. 🧵

[Tom Beer @aTomBeer]

Still, more effort should be put in teaching that the bias variance tradeoff is not a universal truth, at least based on the measures of complexity we use in our everyday practice. This has implications on how we select and evaluate ML models.