rob-hogan.bsky.social
Director of AI @ CergenX, using AI to detect brain injury in newborns // PhD Theoretical Physics | deep learning, science, medtech
https://selfsupervised.substack.com/
43 posts
312 followers
787 following
Regular Contributor
Active Commenter
comment in response to
post
This represents a significant milestone for the field with the first ever demonstration of human expert level performance on held out datasets.
comment in response to
post
First of many for CergenX with our friends in INFANT Research Centre.
Thanks to our co-authors Geraldine Boylan, Sean Griffin, Aurel Luca, Sean Mathieson, and Soraia Ventura.
Supported by Enterprise Ireland DTIF grant
comment in response to
post
😅 definitely part of it. But I think of it more a breakthrough in the crappy post training for prompting issue LLMs have. I wrote a bit more about it here if you're interested open.substack.com/pub/selfsupe...
comment in response to
post
"Our robots are safe, lightweight, and inexpensive - perfect for deploying novel AI algorithms, testing new capabilities among people, and collecting data at scale"
Sounds amazing. Target price point? Would love a hobby robot to train at home.
comment in response to
post
It's an interesting idea to be able to ingest a someones feed and have an LLM judgement of whether they meet some niche category of interest or personality type. Basically automated personal.
comment in response to
post
Interesting, but not sure only ever seeing thing you like is a wise goal. Worrying close to "things I agree with"
comment in response to
post
I'm genuinely struggling to find accounts on here that talk about AI in any way that's not pure hatred. I know X is full of insufferable no nothing shills but still ... a bit weird that a whole technology class is apparently political?
comment in response to
post
It's a small but interesting first step in using AI as a tool for science in this field. We have more promising results under review for early next year.
But we are just scratching the surface and have big plans to dig deeper.
You can checkout full poster here www.cergenx.com/resources/re...
comment in response to
post
With just 1 hour of data we could can predict a significant difference in seizure burden between localised and generalised groups.
This was previously unknown and could ultimately to different treatment paths for these groups.
comment in response to
post
Neonatal EEG is specialized & underexplored at this scale. Our models is the first to perform at level of human expert and can annotate channel-by-channel. This unlock a new view into spatial distribution of seizures in large scale data.
comment in response to
post
How did you arrive at that conclusion? The whole point is to save on compute resources.
comment in response to
post
Hey, could I please be added? scholar.google.com/citations?us...
(Once a physicist, always a physicist)
comment in response to
post
To be clear, I'm Director of AI at a MedTech startup. So clearly believe in AI. But do think there's be an over focus on a solution looking for a problem. Vision was the 1st modality to work.
In our case, the expertise just isn't there so babies are simply doing without the gold standard of care.
comment in response to
post
As an aside, I think ARC is massively overhyped and much more specific than claimed.
Will reserve judgement on TTT when I see it improve on some other tough benchmarks like Frontier Math
comment in response to
post
✋ please and thanks
comment in response to
post
Cautionary tale for anyone trying to rollout AI in MedTech. You need to spend time understanding the work of the end user clinicians.
comment in response to
post
I totally agree this is a problem area for MedAI esp. when datasets aren't public. Also found that people pick and choose metrics to suit themselves. I open sourced an eval library to standardize metrics for my field. I've considered using it as a public leaderboard i.e. open PR w/ predictions
comment in response to
post
Clearly they must be forgetting to cite the real originator of all their ideas
comment in response to
post
😅
Seriously though, that's a pretty good measure for the slow down in publishing by the big labs. Fairly stark given real growth in active researchers
comment in response to
post
✋, please. Thanks
comment in response to
post
Nice work!
I remember being quite surprised by Medprompt results from Microsoft outperforming fine-tuned models with vanilla GPT4.
The prompt sensitivity of these models definitely muddies the waters.
comment in response to
post
✋ please, thanks
comment in response to
post
I came to AI from physics and CergenX I'm really glad to be back in a field where AI Research and scientific discovery are hand in hand. We've only taken our first baby steps in this direction but there is much more to come.