Vivek Natarajan
@vivnat
Followers
5,550
Following
368
Media
58
Statuses
662
Advancing Biomedical AI @GoogleAI | Previously @facebookai Research | Definite Optimist
Mountain View, CA
Joined December 2009
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
بايرن
• 524729 Tweets
موناكو
• 523341 Tweets
Wizkid
• 302438 Tweets
Davido
• 283723 Tweets
Adams
• 124243 Tweets
#BrightestStarBangChan
• 109850 Tweets
#우리의_찬란한_청춘_방찬에게
• 108676 Tweets
Doug Emhoff
• 70881 Tweets
Shakira
• 51548 Tweets
CHAE X JESBIBLE
• 45204 Tweets
Shana Tova
• 41457 Tweets
Centro Carter
• 38258 Tweets
Gago
• 35040 Tweets
KARIME EN HOY
• 34676 Tweets
Cassper
• 34104 Tweets
Rosh Hashanah
• 33218 Tweets
GALA EN HOY
• 32610 Tweets
Lille
• 25515 Tweets
センラさん
• 17156 Tweets
NASCAR
• 16640 Tweets
Girona
• 14784 Tweets
Channie
• 14631 Tweets
Bologna
• 14553 Tweets
Marburg
• 13948 Tweets
ITZY IS THE MOMENT
• 11311 Tweets
Feyenoord
• 10390 Tweets
Camavinga
• 10100 Tweets
Pinned Tweet
Delighted to share our new
@GoogleHealth
@GoogleAI
@Deepmind
paper at the intersection of LLMs + health.
Our LLMs building on Flan-PaLM reach SOTA on multiple medical question answering datasets including 67.6% on MedQA USMLE (+17% over prior work).
37
368
2K
Medicine is inherently multimodal.
Thrilled to share Med-PaLM M, the first demonstration of a generalist multimodal biomedical AI system with a stellar team
@GoogleAI
@GoogleDeepMind
@GoogleHealth
Paper:
12
108
456
Pleased to share our latest
@GoogleAI
@GoogleHealth
pre-print on Med-PaLM 2 where we detail our progress towards physician expert-level medical question answering performance!
Link:
7
83
377
Delighted to share that our Large Language Model (LLM) for medicine, Med-PaLM, has been published in
@Nature
today.
With a dream team of co-authors
@alan_karthi
@AziziShekoofeh
@taotu831
@thekaransinghal
and many more at
@GoogleAI
@GoogleHealth
12
64
336
Democratizing access to world-class medical expertise is my life mission.
Our
@GoogleAI
@GoogleDeepMind
@GoogleHealth
research diagnostic dialogue AI, AMIE, is a tantalizing glimpse of the future with AI at the heart of care.
Google AI blog -
9
29
201
Alan sharing our latest
@GoogleHealth
@GoogleAI
research : our new SOTA medical LLM, Med-PaLM 2
On USMLE MedQA, Med-PaLM 2 reaches an accuracy of over 85% going from ‘passing score’ to expert performance
With Tao Tu,
@Mysiak
,
@thekaransinghal
,
@AziziShekoofeh
and
@alan_karthi
5
35
192
alphafold 3, human brain connectomics, med-gemini, amie, …
i try not to think about competitors too much, but i cannot stop thinking about the aesthetic difference between using ai to solve deep, hard grand challenges in medicine & science and building waifus
🙃
Fun to hear Google SVP James Manyika at IO mention our work on Med-Gemini as one of the biggest opportunities and applications for AI. Exciting things to come 🚀
2
6
52
8
12
162
Fun new
@GoogleAI
@GoogleHealth
@StanfordMed
preprint exploring the potential of LLMs for genetic discovery!
We found that our medical LLM, Med-PaLM 2, was able to accurately identify the murine genes that contained experimentally verified causative genetic factors for six
A baby step towards using LLM for Genetic Discovery. Med-PaLM 2 helps identify murine genes linked to hearing loss! While promising, limitations exist (tokenizer 🔡, quality of pretraining data✅)
#AIforScience
#GeneticDiscovery
@GoogleAI
@StanfordMed
0
11
44
2
21
132
A rather generous headline but very grateful to
@Analyticsindiam
for the opportunity to share more about my research journey and inspirations, our work on Med-PaLM at
@GoogleAI
and my vision and dreams for LLMs and AI in biomedicine.
9
10
126
I got my EB1-A green card last year - it’s life altering.
Indebted to the incredible generosity of folks I met through this platform who wrote recommendations for me.
Hope to pay it forward - if you are going through the process and need letters, please reach out if helpful.
4
1
82
In addition to objective metrics, we pilot a framework for clinician/layperson eval of answers revealing key gaps in Flan-PaLM responses.
To resolve this, we use instruction prompting tuning to further align LLMs to the medical domain and generate safe, helpful answers.
1
4
68
However, the Med-PaLM answers remain inferior to clinicians overall, suggesting further research is necessary before LLMs become viable for clinical applications.
We look forward to careful and responsible innovation to drive further progress in this safety-critical domain.
4
13
66
Med-PaLM performs encouragingly on several axes such as scientific and clinical precision, reading comprehension, recall of medical knowledge, medical reasoning and utility compared to Flan-PaLM.
4
6
61
Work done with an amazing, interdisciplinary team of co-authors including
@thekaransinghal
,
@AziziShekoofeh
, Tao Tu,
@alan_karthi
, Sara Mahdavi Chris Semturs,
@_jasonwei
,
@alvin_rajkomar
,
@weballergy
,
@martin_sen
,
@chrisck
,
@stephenpfohl
and many more.
4
1
55
When people ask about Med-PaLM 2 vs GPT-4, emails such as these are what we point to :)
3
5
54
I am really looking forward to sharing our work on AMIE and LLMs for biomedicine at this
@GoogleIndia
Research@ event tomorrow in Bangalore along with many other amazing speakers!
If you are around, please come say hi!
I'm looking forward to attending and speaking at the Research@ Bangalore event tomorrow!
@GoogleIndia
11
40
577
0
3
53
Think this is my favorite vision paper since ... Vision Transformers.
Super creative work again from this stellar team!
New paper from Brain Zurich and Berlin!
We try a conv and attention free vision architecture: MLP-Mixer ()
Simple is good, so we went as minimalist as possible (just MLPs!) to see whether modern training methods & data is sufficient...
18
218
989
0
3
45
Exciting transfer learning work from team mates in the
@GoogleHealth
Radiology team with the release of CXR Foundation - a tool that can significantly reduce data needed to train chest x-ray AI models and accelerate safe medical AI development.
A new
#GoogleAI
tool reduces the size of datasets needed for radiology models used to predict abnormalities in chest x-rays, making it easier for researchers to build custom models that can help in disease detection.
73
202
1K
3
6
43
Delighted to share this preview of the multimodal capabilities we are bringing to Med-PaLM!
Medicine is inherently multimodal.
These advancements are a leap towards generalist biomedical AI that can integrate all this information at scale and power exciting new applications!
Med-PaLM 2 can help answer questions and summarize insights from a variety of dense medical texts. We’re working to add more capabilities to Med-PaLM 2, so that it can synthesize information from medical imaging like mammograms or help radiologists interpret results.
#GoogleIO
7
41
194
4
4
41
@emollick
Thanks so much for highlighting our work!
Please see thread below for more context on the work and especially its limitations.
Delighted to share our new
@GoogleHealth
@GoogleAI
@Deepmind
paper at the intersection of LLMs + health.
Our LLMs building on Flan-PaLM reach SOTA on multiple medical question answering datasets including 67.6% on MedQA USMLE (+17% over prior work).
37
368
2K
1
5
41
@HarshaNori
Hi Harsha, thanks for the comments!
Our x-axis in the Med-PaLM 2 paper is upto Mar 23. We announced our Med-PaLM 2 results at our Health event on Mar 14, 23.
At that time, our results on MedQA (~85%) far surpassed other AI systems including Med-PaLM (~67%) and GPT-3.5 (~60%).
2
4
40
Med-PaLM M is a large multimodal generative model that flexibly encodes and interprets biomedical data spanning clinical language, medical imaging, genomics and more performing competently on a diverse array of tasks - all with the same set of model weights.
1
6
34
Genuinely humbled to be on this panel on LLMs in healthcare
@SAILhealth
today.
Folks on the panel and in the audience are luminaries who have inspired me to do what I am doing today.
Grateful to
@zakkohane
@arjunmanrai
for the opportunity.
Enjoying listening to
@SebastienBubeck
@srikanthbelwadi
@vivnat
on a panel about LLMs in healthcare
#SAIL23
. Sébastien shows how drawing pictures of unicorns is related to the future of medicine.
3
8
39
1
3
33
You will be working with a diverse team of researchers and clinicians including
@thekaransinghal
,
@AziziShekoofeh
,
@JanFreyberg
and
@alan_karthi
.
If interested, please reach out!
13
1
33
Self-supervised learning has rapidly played a key role in helping us build more performant, robust and label efficient medical imaging models
@GoogleHealth
.
Delighted to be able to able to share some of the progress led by the exceptional
@AziziShekoofeh
Medical image classification models often pre-train on natural image datasets. Today, we present alternative approaches that use additional pre-training on medical images, along with metadata-based data augmentation, to significantly improve performance.
5
152
537
1
5
32
With transformers catalyzing progress, its an exciting time to be working in biomedical AI with amazing possibilities for the advancement of science and human health!
Thanks to
@stevenyfeng
,
@RylanSchaeffer
and
@DivGarg9
for having me. Truly delightful experience
@Stanford
!
Week seven of CS25 - Transformers United: We had
@vivnat
from
@GoogleAI
,
@GoogleHealth
give the first public talk on Med-PaLM: the current SOTA Transformer model for Medical Q/A & on building Foundation Models for Medical Science and Healthcare
A very interesting talk
5
3
72
2
3
31
Our new work "Addressing the real world class imbalance problem in dermatology"-
has been accepted as a full paper at
#NeurIPS2020
ML4H workshop!
Work led by brilliant
@ckbjimmy
with
@GoogleHealth
/Brain team mates Jon Deaton,
@gamaleldinfe
and Yuan Liu
1
5
32
AMIE :)
Geoffrey Hinton says AI doctors who have seen 100 million patients will be much better than human doctors and able to diagnose rare conditions more accurately
44
137
610
0
3
32
The RETFound
@Nature
paper is a milestone for Medical AI - extremely impressive performance across a suite of retinal imaging and oculomics tasks. Model also open source!
Hearty congratulations
@Yukunzhou19
@markachia
@sktywagner
@pearsekeane
and team!!
1/
🚨🚨 New paper alert 🚨🚨
Introducing RETFound, a foundation model for ophthalmology
We’re super excited about this and hope it will act as a
#Cornerstone
for global efforts to prevent blindness through
#AI
@UCLeye
@Moorfields
#OpenAccess
@Nature
47
261
714
2
5
29
Thread contextualizing Med-PaLM 2 v GPT-4 results reported in our paper.
Benchmarks such as MedQA (USMLE) are nice but misleading indicators of LLM performance and rigorous evaluation is necessary.
Our work is fully focussed on that and we encourage the community to do the same
@HarshaNori
Hi Harsha, thanks for the comments!
Our x-axis in the Med-PaLM 2 paper is upto Mar 23. We announced our Med-PaLM 2 results at our Health event on Mar 14, 23.
At that time, our results on MedQA (~85%) far surpassed other AI systems including Med-PaLM (~67%) and GPT-3.5 (~60%).
2
4
40
2
7
29
One of those moments when the emotion is so overwhelming that there no words left, only tears.
Thank you team India 🏏🙏🏆
2
1
28
"If this was to be made available to every X-ray center in rural India, I think we could beat TB"
Heartwarming to see AI for medical imaging reach remote Indian districts like Simdega and help fight deadly diseases like TB (by
@apoorva_nyc
)
3
5
26
@IAmSamFin
@seb_ruder
A few ML ones I like
- The Batch from Andrew Ng et al -
- Import AI from Jack Clark -
- Skynet today's
For AI and medicine, there is
0
6
25
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
MutliMedBench spans 7 biomedical modalities and 14 diverse tasks such as medical QA, radiology report generation, genomic variant calling and more with over 1 milliion samples.
We hope the curation of MultiMedBench will further spur the development of generalist biomedical AI.
1
5
24
Nice work improving on EfficientNets but
- All lambda (L) layers has a ~7x reduction in throughput over CNNs
- Best model wrt speed-accuracy tradeoff has only a few L layers in c4 stage of Resnet
Predict future will be such hybrids but hard to see CNNs go away completely
Lambda Networks: SOTA on ImageNet. Again Transformer-like architectures dominating over a new field 🚀. CNNs had a good run 😰.
Paper:
Awesome video from
@ykilcher
:
1
110
556
1
9
24
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
Med-PaLM M is built by fine tuning and aligning PaLM-E - an embodied multimodal language model from
@GoogleAI
- to the biomedical domain using MultiMedBench, a newly curated open source biomedical benchmark.
1
0
21
Incredibly excited to see many years of hard work from the team coming to fruition.
One important step closer in taking AI in dermatology from research to the real world!
2 billion people globally suffer from skin conditions. Our CE-marked AI-powered dermatology tool will help people research common skin conditions.
#GoogleIO
9
168
569
0
1
20
Important paper from Percy Liang's group showing how selective classification intended to reduce costly real world AI errors by abstention can instead end up magnifying disparities between subgroups especially in the presence of spurious correlations.
1
6
20
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
Across all tasks in MultiMedBench, Med-PaLM M reaches performance competitive or exceeding SOTA, often exceeding specialist models by a wide margin.
Further, Med-PaLM M also significantly outperforms PaLM-E demonstrating the importance of biomedical fine tuning and alignment.
1
1
18
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
For eg, we find that Med-PaLM M can accurately identify and describe TB in chest x-rays despite having never encountered presentations of the disease before in images - only through language based instructions and prompts.
1
2
17
Our
@GoogleHealth
dermatology team's work on "A Deep Learning System for the Differential Diagnosis of Skin Diseases" featured in the cover of
@NatureMedicine
's June issue!
We are excited to announce that our paper, “A Deep Learning System for Differential Diagnosis of Skin Diseases", has been published in
@NatureMedicine
. You can see the paper and learn more about this research in our blog post! ↓
13
245
906
2
3
18
Super excited about giving a talk tomorrow at
@MedTechatUCL
's 'AI in Medicine' series on
'Building better medical imaging AI for clinical deployment at scale'
Event details -
@UCL
AI in Medicine Series continues next week bringing together some of the greatest minds in digital health. Don't miss out! Get your tickets now to see
@manoliskellis
,
@EricTopol
,
@andreascleve
,
@siddhartha_c
,
@vivnat
1
3
14
1
2
17
Excited to share our latest
@GoogleHealth
-
@DeepMind
collaboration!
Safety and reliability are key challenges to deploying AI at scale in real world clinics and reliable out of distribution (OOD) detection is an important problem in this space.
New paper! Joint contrastive and supervised training improves OOD detection performance on the challenging near OOD setting by obtaining a rich and task-agnostic feature space.
Thread.
1
19
103
1
3
17
as for the Alfred Nobel reference on the board, we ended the class speculating in which field AI might win the first Nobel Prize.
My bet is on medicine :)
2
0
14
This is fantastic work! Highly recommend reading this if you are working in ML for Health.
New perspective in
@Nature_NPJ
Digital Medicine w/
@latentjasper
and
@AndrewLBeam
!
"Second opinion needed: communicating uncertainty in medical machine learning"
🧵 below! 1/n
6
43
158
1
2
15
Incredibly fortunate enough to have witnessed the 2001
#AUSvIND
series and the Eden Gardens miracle but this series win has to be the greatest in Indian test cricket history.
#Pant
#Rahane
#Shubmangill
1
0
14
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
@GoogleHealth
@GoogleDeepMind
Huge props to
@taotu831
for driving this work with relentless energy and optimism together with stellar teammates
@AziziShekoofeh
@DannyDriess
@HardyShakerman
@peteflorence
@thekaransinghal
@alan_karthi
@pichuan
across
@GoogleAI
@GoogleDeepMind
@GoogleHealth
(1/2)
1
1
14
Looking forward to speaking here and meeting the AI LA community working on health and life sciences!
Keynote Announcement 👀 👀
We’re kicking off our
#LifeSummit22
on Oct 20th @ 11:00 AM PST with
@vivnat
Health AI Researcher at
@Google
to talk about today’s landscape of AI in Healthcare and what we can anticipate in 2023 and beyond!
RSVP 👉
0
2
5
0
1
14
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
Towards that, we find preliminary but exciting evidence that Med-PaLM M can generalize to novel medical tasks and concepts and perform zero-shot multimodal reasoning - all in a zero-shot fashion only through language based instructions and prompts.
1
0
14
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
A key intuition for building large scale generalist biomedical AI with language as a common grounding across tasks is the possibility of combinatorial generalization and positive task transfer.
1
0
13
Perhaps, most excitingly, in side by side comparison with physician generated answers, Med-PaLM 2’s answers were preferred across eight of nine axes relevant to clinical utility, such as factuality, medical reasoning capability, and low likelihood of harm.
1
0
13
This pocket sized book is likely going to be my most profound read ever. Bestows so much serenity and calmness on the reader.
1
1
13
We also have SOTA results on other benchmarks such as MedMCQA and MMLU clinical topics by a large margin.
More details below:
Preprint on Med-PaLM 2 out soon :)
2
2
13
With a stellar team
@GoogleAI
@GoogleHealth
@DeepMind
including
@thekaransinghal
,
@AziziShekoofeh
,
@taotu831
,
@alan_karthi
,
@Mysiak
, Rory Sayres, Ellery Wulczyn,
@stephenpfohl
,
@weballergy
,
@greg_corrado
,
@ymatias
and many others.
1
0
12
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
Our paper has more experiments, details and insights.
Would love your feedback -
Zooming out, generalist biomedical AI is a dream many of us at
@GoogleAI
@GoogleHealth
@GoogleDeepMind
have been building towards for years.
1
0
12
V. excited about this work expertly led by
@abzz4ssj
&
@jessierenjie
+ team where we use a novel yet simple hierarchical outlier detection loss + ensemble diversity to tackle long tail OOD detection (a most important problem in medical AI safety) for dermatology.
Our new paper tackles an important safety hurdle for ML from code to clinic- “how does your dermatology classifier know what it doesn’t know?” In clinical practice patients may present with conditions unseen by ML systems in training, causing errors 1/2
2
18
77
1
1
12
@anshulkundaje
Anshul, thanks for the feedback. We believe in openly sharing progress but also in safe + responsible innovation.
We have been working with many academic + industry partners to evaluate Med-PaLM including from Stanford.
If you would like to collaborate, feel free to reach out.
2
1
12
Pant channeling peak inner Dhoni. Sorry
@SriniMaama16
but Pant over DK surely now for the World Cup.
2
1
11
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
@GoogleHealth
@GoogleDeepMind
The possibilities with such generalist biomedical AI that can encode the biomedical universe is limitless with applications spanning the continuum of scientific biomedical discovery to care delivery.
The future of AI in medicine and bio is incredibly exciting!
1
1
12
some qualitative examples here below comparing Med-PaLM 2 and Med-PaLM responses.
1
1
11
in collaboration with brilliant team mates
@mo_norouzi
,
@_basilM
,
@JanFreyberg
,
@alan_karthi
,
@tingchenai
,
@skornblith
,
@JonJonDeats
,
@fionakryan
, Zach Beaver & Aaron Loh across
@GoogleHealth
and Google Brain.
0
0
11
This hard hitting book by
@MartyMakary
is the best book I have read this year. If you work in healthcare, this is an absolute must read. I have huge personal reasons for working in health but this has supercharged me. h/t to
@rabois
for the recommendation.
1
0
11
Hearty congratulations!!!
Working with Alan daily is one of the best parts of my job.
Alan’s enthusiasm, passion and vision to make healthcare better for everyone is an incredible force multiplier and drives the team forward.
Look forward to everything that’s coming :)
Honour to be included among this list of innovators in digital health- thank you
@IntHealthAI
💙
12
1
63
1
0
11
Today we celebrated our dad’s ‘official’ birthday (even though he isn’t the biggest fan of such celebrations).
2
0
11
New work from team mates
@GoogleHealth
dermatology team published in JAMA Network Open showing PCPs and NPs (~20% more accurate) can potentially provide better dermatologic care to patients using AI assistance. Suggests 1 in 10 patients would benefit.
0
2
10
@suchisaria
Agree with the article. Think the research community needs to appreciate challenges of applied work more. Far too much focus on novelty in conferences reviews but the sad fact is almost all of these 'novel' methods are not scalable/applicable to real world production systems.
1
0
10
Devi/Dhruv are not only brilliant scientists but also among the warmest/kindest people I have known.
In fact, how I approach work/life today is hugely influenced by observing them at close quarters.
Glad others can also get to learn from them through this refreshing interview.
Episode 1 is out! Dhruv Batra (
@DhruvBatraDB
) on Humans of AI: Stories, Not Stats.
Video:
Podcast:
All episodes so far:
10
39
364
1
0
10
Exciting new work using AI to improve colonoscopy from teammates at
@GoogleHealth
!
Announcing C2D2, a
#MachineLearning
-based approach to improving colonoscopy screening coverage that performs real-time local 3D reconstruction of the colon during the procedure, and identifies regions outside the field of view of the endoscope. Read more ↓
6
146
387
1
1
10
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
@GoogleHealth
@GoogleDeepMind
@pichuan
Mohamed Amin, Sara Mahdavi, Chris Semturs, Joelle Barral,
@skornblith
@acarroll_ATG
@_basilM
@RyutaroTanno
@achowdhery
@greg_corrado
@ymatias
and many more we have the privilege of working with (2/2)
0
1
10
This excellent work from the Genomics team
@GoogleHealth
is a great example of how much impact we can have on important problems like genome sequencing when state of the art deep learning methods like transformers are combined with great domain knowledge and expertise!
We’re excited to release DeepConsensus, a transformer approach for PacBio HiFi consensus generation. DeepConsensus reduces HiFi read errors by 42%, improving assembly NG50, base accuracy, as well as variant calling. Led by
@gunjan_baid
and
@daniel_e_cook
4
115
270
0
1
10
Med-PaLM 2 builds on top of PaLM 2, the latest SOTA LLM from
@GoogleAI
We further finetune / align the model to the requirements of the medical domain and introduce ensemble refinement as a simple new prompting strategy to improve the model’s medical reasoning capabilities.
1
0
10
Super excited about this work!
Underspecification has important implications when building AI models for healthcare and this paper has several examples from medical imaging, EHR, genomics and even epidemiology to illustrate this.
Please give it a read!
NEW from a big collaboration at Google: Underspecification Presents Challenges for Credibility in Modern Machine Learning
Explores a common failure mode when applying ML to real-world problems. 🧵 1/14
16
258
1K
1
1
9
We also introduce a new adversarial questions benchmark dataset designed to elicit model answers with potential for harm and bias and probe the safety and limitations of LLMs in the medical domain.
1
0
9
Excellent overview from Nathan/Ian. If you work in AI or even just have a passing interest, well worth your time.
Glad to see both my current team's (
@GoogleHealth
) research breakthroughs in AI for ophthalmology and mammography and work with previous team on VQA (LORRA) featured
💥We're live with
@stateofaireport
2020💥
For the 3rd year,
@soundboy
and I spent 2 months analyzing the most interesting developments in AI.
We aim to trigger an informed conversation about the
#stateofai
and its implication for the future. Thread 👇
21
309
752
1
4
9
@sktywagner
@sktywagner
@pearsekeane
Thank you so much for having me today! It was really nice to meet everyone, learn about some really cool research and discuss these interesting topics :) Thoroughly enjoyed it!
0
0
9
@taotu831
@AziziShekoofeh
@HardyShakerman
@peteflorence
@DannyDriess
@thekaransinghal
@alan_karthi
@GoogleAI
Finally, to understand the clinical applicability of Med-PaLM M, we conducted a radiologist evaluation of AI generated reports across model scales.
Clinically significant error rate for Med-PaLM M on par with radiologists from prior studies suggesting potential clinical utility.
2
0
9
Happy Fathers Day, appa! Thank you for showing us how to lead a kind, purposeful and meaningful life and for always being our guiding light. I wish things were better but I know no matter what you will always be our biggest supporter and believer.
Today we celebrated our dad’s ‘official’ birthday (even though he isn’t the biggest fan of such celebrations).
2
0
11
0
0
9
Paper has many insights but one that was super interesting was the significant average predictive disagreement & complementarity we observed in representations trained using supervised & self supervised pre-training.
Could be an exciting direction of research for the community!
0
1
9
(don't want to pile on today but needs saying that you couldn't see a starker difference in culture, priorities and maturity of team / leadership)
1
0
9
@le_roux_nicolas
This depends on the research topic no?
Lightweight reviews are common if you do general research on open data but if work touches nuanced areas, the process is strict.
Context: I work in health and end up spending considerable time shepherding papers through internal reviews.
0
0
9
This has been a riveting read thus far from
@charlesgraeber
. Immunotherapy has a number of interesting parallels to AI going through cycles of hype and bust and rebirth and the single minded dedication of researchers to keep going despite all odds.
1
2
7
A real privilege to working on this mission with an amazing team including
@alan_karthi
@taotu831
@apalepu13
@HardyShakerman
@KhaledSaab11
@JanFreyberg
@RyutaroTanno
@thekaransinghal
@AziziShekoofeh
@weballergy
@Mysiak
@greg_corrado
@ymatias
and many more not active on Twitter!
1
1
8
Work led by the brilliant
@GhorbaniAmirata
last summer on DermGAN: generation of synthetic skin images with pathology featured in today's
@GoogleAI
blog.
Paper -
(was a full paper at ML4H NeurIPS 2019 with
@dav_cz
and Yuan Liu)
Check out two ways of improving the diversity of training data for
#MachineLearning
models in health applications: a configurable method for synthetic skin lesion images of rare skin types and conditions, and synthetic images of defocused tissue samples.
3
112
187
0
1
8
Med-PaLM 2 approaches or exceeds SOTA on multiple medical question answering benchmarks.
In particular, it reaches a score of 86.5% on MEDQA (USMLE) dataset matching performance of expert test takers (>18% over Med-PaLM).
It also reaches a new SOTA of 81.8% on PubMedQA.
1
0
8
A great list!
Will also add two papers from D. Sculley et al to better understand nuances/challenges of ML in production at scale
ML: The high interest credit card of technical debt -
Hidden Technical Debt in ML Systems -
0
3
8
@pdhsu
@taotu831
@KhaledSaab11
and our team previewed related long context Gemini capabilities as part of our Med-Gemini paper - (page 28).
We’d be happy to work with you to improve and further specialize for biomedical research.
1
0
7
@googlecloud
The future is exciting and we have more to share soon!
Grateful to the reviewers / editors and a stellar set of team mates (past and present) at
@GoogleAI
- Chris Semturs, Renee Wong, Sara Mahdavi, Heather Lewis, Yun Liu, Joelle Barral, Dale Webster,
@stephenpfohl
,
@weballergy
,
3
0
8
Extremely excited about this work!
Best part: BiT models freely available here courtesy
@__kolesnikov__
@giffmana
@neilhoulsby
et al and provide a simple way to significantly ⬆️ performance, data efficiency and robustness of your medical imaging models.
New paper from our team
@GoogleHealth
/
@GoogleAI
() Pre-training at scale improves AI accuracy, generalisation + fairness in many medical imaging tasks: Chest X-Ray, Dermatology & Mammography! Led by
@_basilM
,
@JanFreyberg
, Aaron Loh,
@neilhoulsby
,
@vivnat
1
20
97
1
2
8
While further studies are necessary to validate the efficacy of these LLMs in real-world clinical settings, these results nevertheless highlight the rapid progress we are making towards expert-level performance in medical question answering.
The possibilities are exciting!
1
0
8
My facial expression may bely this, but the panel at
@TEDAI2023
today was thoroughly enjoyable!
Today’s panel discussion on
#generativeAI
in healthcare was moderated by
@daniel_kraft
Founder & Chair,
@NextMedHealth
&
@DigitalDHealth
. Huge thank you to our panelists ⬇️ for sharing their insights and predictions about
#AI
’s implications for providers, payers, and patients.
1
2
9
1
0
8
2
0
8
The team at
@GoogleHealth
is amazing! So fortunate to work closely with all the amazing folks here.
Interrupting the
#COVID19
-casting for 2 new and impressive
#AI
in medicine papers
@NatureMedicine
#deeplearning
Skin diseases (26 conditions)
@dav_cz
@google
Yuan Liu
Age-related macular degeneration
@JeffreyDeFauw
@PearseKeane
0
39
80
1
0
8