Adarsh Subbaswamy

Using AI to improve and accelerate medical product development | Assistant Professor @UMBaltimore | CS PhD from @JohnsHopkins

[Adarsh Subbaswamy @_asubbaswamy]

New preprint w/ (co-first author) @royjamesadams and @suchisaria: "Evaluating Models Robustness Under Dataset Shift" https://t.co/ykOu7G9EQE

How can we evaluate *ahead of time* whether or not a model's performance will generalize from training to deployment? 1/

[Jean Feng @Jean_J_Feng]

Our work is finally published in @npjDigitalMed! 🚀 In https://t.co/QAB9D3UJoc , we show that AI is capable of scaling up analyses of medical device regulatory documents, and has a lot to offer to patients, regulators, researchers, and vendors

[Adarsh Subbaswamy @_asubbaswamy]

@roydanroy Wow, huge win for DeepMind! My first ML conference was UAI 2017 in Sydney. I’ll never forget your talk on the “nonvacuous generalization bounds” paper. Congrats!

[ARPA-H @ARPA_H]

What if a platform existed to help doctors make faster more accurate diagnoses?

INDEX intends to create a high-quality medical imaging platform that does not currently exist to benefit the entire imaging ecosystem. https://t.co/IVfovONQiA

[Adarsh Subbaswamy @_asubbaswamy]

@beenwrekt @srush_nlp Do we have good answers for this at the grad or research level? 😅

[Adarsh Subbaswamy @_asubbaswamy]

@deliprao @jeremyjkun I like the music analogy. I wonder if it’s like guitar tabs vs sheet music

[Adarsh Subbaswamy @_asubbaswamy]

@beenwrekt @jjcherian @iswgibbs If you define a set of groups with respect to covariate values (e.g., demographic info), then if you shift the prevalence of the groups you will see a corresponding covariate shift . Not sure if the reverse direction works the same

[Adarsh Subbaswamy @_asubbaswamy]

@aahladpuli 👀I discovered the stormlight archive this year

[Adarsh Subbaswamy @_asubbaswamy]

@tallinzen @jbhuang0604 Referencing this I imagine

https://t.co/99mie1ZMaW

[Adarsh Subbaswamy @_asubbaswamy]

It turns out monitoring the performance of ML models is hard, especially in settings like healthcare where model predictions affect user decisions

If you're at NeurIPS check out Jean's talk and we'd love to hear your thoughts re monitoring

[Jean Feng @Jean_J_Feng]

Our framework brings together ideas from causal inference and statistical process control. Please drop by if you're interested! Preprint here: https://t.co/ROvBdBudOy It was wonderful working on this with coauthors from UCSF and FDA! @pirracchior @_asubbaswamy @harvineet_singh

[Jean Feng @Jean_J_Feng]

Our framework brings together ideas from causal inference and statistical process control. Please drop by if you're interested! Preprint here: https://t.co/ROvBdBudOy It was wonderful working on this with coauthors from UCSF and FDA! @pirracchior @_asubbaswamy @harvineet_singh

[Jean Feng @Jean_J_Feng]

I'm talking at today's NeurIPS #RegulatableML workshop about why designing monitoring systems for ML-based medical devices is complicated, why we need a systematic framework for comparing all the different monitoring strategies, and our first steps to developing such a framework.

[Suresh Venkatasubramanian (mostly in the sky now) @geomblog]

How do people manage to keep track of ML papers? This is not a request for support in my current state of bewilderment - I'm genuinely asking what strategies seem to work to read (or "read") what appear to be 100s of papers per day.

[Adarsh Subbaswamy @_asubbaswamy]

@vitaliikl @St_Megas @rlmcelreath Assuming faithfulness, the collider is distinguishable from the pipe and fork. As the image states, in the collider case X and Y are marginally independent while for the other two X and Y are marginally associated. See, e.g., https://t.co/DaAaUNJRaN

[Adarsh Subbaswamy @_asubbaswamy]

@sjmielke Congrats!

[Adarsh Subbaswamy @_asubbaswamy]

@ccanonne_ I think the fish really help make this math more digestible

[Adarsh Subbaswamy @_asubbaswamy]

@roydanroy I'm unfortunately having to do this... I'm about to watch the best paper in my batch get rejected because of a spiteful reviewer and uninformed AC, and the worst paper get accepted because of uninformed reviewers and uninvolved AC

Why do people care about neurips papers again?

[Adarsh Subbaswamy @_asubbaswamy]

@EugeneVinitsky @ccanonne_ @shortstein imo this only really happens when papers are shepherded and there are opportunities for revisions. I’ve had much better experiences as an author with journal submissions because of this

[Yoav Wald @wald_yoav]

Our #UAI2023 paper is on detecting novel subgroups (aka classes/categories) when distribution of non-novel data shifts.
If this sounds interesting, or you’re wondering what it has to do with unicorns, check out our poster/paper/drop a line.

w/ @suchisaria
https://t.co/7nrMfDblfx

[Adarsh Subbaswamy @_asubbaswamy]

@ElanRosenfeld Has the avg number of reviews per paper gone up over time? I remember 3 per paper being common...

[Suchi Saria @suchisaria]

If you’re at #ICML2023 join me here: https://t.co/Kzca5T5dCf

Giving an invited talk later today on AI Safety.

Below are some of the papers our lab has written on the topic in last 5 yrs; this talk will give a brief overview and focus on learning/evaluation.