Roopal Garg
@roopalgarg
Followers
684
Following
934
Media
36
Statuses
853
Staff Software Engineer @GoogleDeepMind | Gemini | @GoogleAI l Multimodality | Multi/Cross linguality | @USCViterbi @CSatUSC alumnus | Views are my own
Austin, TX
Joined March 2010
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Leyen
• 281762 Tweets
梅雨明け
• 216960 Tweets
#BKPPยอม_MV
• 197379 Tweets
norawit x bv in emporium
• 115084 Tweets
INTRO
• 78744 Tweets
Papa Dali
• 54762 Tweets
Ali Koç
• 52834 Tweets
Kamari
• 51663 Tweets
パワプロ
• 48970 Tweets
MEW IN FRED SEIBU
• 48881 Tweets
岸田総理
• 36067 Tweets
VLOG 5 WITH BUILD
• 33358 Tweets
#素のまんま
• 33238 Tweets
डिब्रूगढ़ एक्सप्रेस
• 29433 Tweets
Mehmet Büyükekşi
• 26357 Tweets
रेल मंत्री
• 23634 Tweets
#TrainAccident
• 23164 Tweets
AMIA
• 21621 Tweets
国立公園
• 17187 Tweets
マンション爆発
• 16897 Tweets
高級ホテル誘致
• 14427 Tweets
What's your color
• 13521 Tweets
#THE夜会
• 13277 Tweets
高級リゾートホテル誘致
• 13143 Tweets
#ミュージックジェネレーション
• 11945 Tweets
西川口の爆発
• 11804 Tweets
Kimiko
• 10799 Tweets
Pinned Tweet
📢 Excited to unveil our latest research, ImageInWords (IIW)! 🚀We're pushing the boundaries of image descriptions with a new seeded, sequential, human-in-the-loop approach producing SOTA, articulate, hyper-detailed descriptions.
arXiv: 🧵1/12
5
32
127
I’m happy to share that I’m transitioning to
@GoogleDeepMind
and continuing work on multimodal multilingual cross-cultural content understanding for Gemini !
Excited about the journey, collaborations and progress we make...
#Google
#GoogleDeepmind
1
1
109
@Gradio
@huggingface
thank you
@Gradio
for sharing our project! Huge thanks for the demo functionality support in
@huggingface
for reference, full 🧵
📢 Excited to unveil our latest research, ImageInWords (IIW)! 🚀We're pushing the boundaries of image descriptions with a new seeded, sequential, human-in-the-loop approach producing SOTA, articulate, hyper-detailed descriptions.
arXiv: 🧵1/12
5
32
127
0
6
17
0
0
15
0
4
9
@arankomatsuzaki
thank you
@arankomatsuzaki
for sharing our work!
Full post 🧵
📢 Excited to unveil our latest research, ImageInWords (IIW)! 🚀We're pushing the boundaries of image descriptions with a new seeded, sequential, human-in-the-loop approach producing SOTA, articulate, hyper-detailed descriptions.
arXiv: 🧵1/12
5
32
127
0
2
10
0
1
8
Congratulations to the IIW Team on the collaboration:
@roopalgarg
@AndreaDBurns
@bka3
@YonatanBitton
@CesleeDineen
@yasumasa_onoe
@andrewbunner
@RanjayKrishna
@jasonbaldridge
@RSoricut
Project Page:
🧵2/12
1
0
8
A weekend of
#100DaysOfCode
awaits!!! What have u planned up?? Working on a library to make training models, adding new models, monitoring and versioning your experiments super easy using Tensorflow! Update to follow soon. Stay tuned.
#DeepLearning
#MachineLearning
#TensorFlow
0
6
8
We release a subset of IIW human- and model-annotated descriptions, as well as human SxS results on Human-Human and Model-Human sourced pairs of descriptions.
Available on:
GitHub:
Hugging Face:
🧵10/12
1
0
6
Dive into the details of our carefully designed seeded, sequential, human-in-the-loop annotation framework which builds on top of existing VLM functionality and uses granular object level human-authored descriptions to compose overall hyper-detailed image descriptions. 🧵3/12
1
0
7
@gowthami_s
my 2c would be to not base your metric on a review from 3 random (perhaps not expert themselves) people's understanding of your work.
if you believe in it and it has true merit.. it will get rewarded..either directly or becomes a base for future success! all the best
1
0
7
#activelearning
is next on my mind. Training a model on static one time data is one thing. making sure its performance doesn't deteriorate with shift in data 6 months down the line is another.
#MachineLearning
#DeepLearning
#datacollection
#NLProc
0
1
7
after 3 years with
@GumGumAILabs
, its time to move on. Its been a great experience and a huge learning curve, want to thank and wish
@GumGum
luck.
I would be joining Google Core Search team in MTV starting next week. Really excited and pumped up for it.
#GoogleSearch
#Google
0
1
6
useful read: Axiomatic Attribution for Deep Networks
For those interested in interpretability of NN
Day 1&2 of
#100DaysOfMLCode
#DeepLearning
#NeuralNetworks
#interpretability
#attribution
0
2
6
accepted!
@sirajraval
hope reading ML/DL/NLP papers counts... i can share reviews, summaries and code insights for implementing and working with the papers
#100DaysOfMLCode
#MachineLearning
#DeepLearning
Who’s ready to take the 100 days of ML code challenge? That means coding machine learning for at least an hour everyday for the next 100 days. Pledge with the
#100DaysOfMLCode
hashtag, I’ll give the first few winners a shoutout!
265
159
957
0
0
6
Focusing on applications, we first evaluate the value of the VLM fine-tuned models on a T2I use-case by comparing the image reconstruction capability of a T2I model on inputs from the VLM fine-tuned models and demonstrate that IIW outputs come out consistently on the top. 🧵8/12
1
0
5
Slides from my
@BigDataDayLA
talk on Deep Learning for Natural Language Processing
#DeepLearning
#NLP
#LSTM
#RNN
0
6
4
My second belong:
Deep Learning for NLP : Part 2
@GumGum
#NLProc
#DeepLearning
#RNN
#LSTM
#GRU
#NeuralNetworks
#AI
#DL
Feedback welcomed
1
1
4
glad to see ImageInWords () still trends on
@huggingface
datasets () after 3 weeks of release!
looking forward to how the community learns and benefits from it!
Stay tuned for future updates and releases.
🧵:
📢 Excited to unveil our latest research, ImageInWords (IIW)! 🚀We're pushing the boundaries of image descriptions with a new seeded, sequential, human-in-the-loop approach producing SOTA, articulate, hyper-detailed descriptions.
arXiv: 🧵1/12
5
32
127
0
1
5
one stop solution to learn everything
#AI
@sirajraval
#DeepLearning
#NeuralNetworks
#TensorFlow
#Blockchain
0
6
3
With the release of our IIW detailed annotation guideline, our goal is to seek community feedback and iteratively make them holistic, reduce human effort and dependency in the annotation process, and help shift the narrative from captions to descriptions.
🧵11/12
1
0
3
Nostalgic presenting in a room not too long ago wrote an exam in...
#BigDataDayLA
@BigDataDayLA
@GumGum
#DeepLearning
#NLP
0
1
4
excellent article... “Highway Networks with TensorFlow” by
@jimmfleming
#DeepLearning
#NeuralNetworks
0
1
4
@kristileilani
Thank you
@kristileilani
for sharing our work 🙏!!
For full 🧵
📢 Excited to unveil our latest research, ImageInWords (IIW)! 🚀We're pushing the boundaries of image descriptions with a new seeded, sequential, human-in-the-loop approach producing SOTA, articulate, hyper-detailed descriptions.
arXiv: 🧵1/12
5
32
127
0
0
1
IIW human-authored dataset achieves a new best-of-the-breed SOTA 🏆 status when compared to comparable prior works on a suite of automated readability based metrics and further validated through in-depth human-SxS analysis across multiple axes. 🧵6/12
1
0
3
The sequential annotation setup induces a human learning environment which results in higher annotation quality in lesser time. Compared to a sequential approach, we get an all-inclusive output verified and augmented by multiple humans with +20% words in -30% time. 🧵4/12
1
0
3
IIW fine-tuned PaLI-5B achieves considerable gains over fine-tuned PaLI-5B from comparable work. When compared to considerably larger models or human-annotations, the model size can be shown as a limitation. We aim to explore this further with iterations on larger models. 🧵7/12
1
0
3
We evaluate across automated and meticulously designed Human-SxS to focus on richness, readability, comprehensiveness, specificity, hallucinations, and human-likeness. The evaluations are run to evaluate both IIW human-authored and downstream finetuned VLM model outputs. 🧵5/12
1
0
3
Learning
#Theano
,
#TensorFlow
for building
#neuralnetwork
for
#NLP
based tasks. Any suggestions what works better and when?
#deeplearning
0
0
3
First shot at writing a post. Feedback welcomed
#DeepLearning
#NLP
#wordembedding
#AI
#DL
#NeuralNetworks
@GumGum
0
1
3
1
0
3
@amuellerml
Averaging out the word embeddings for words on their Wikipedia page would be the simplest and a decent first step. Worked before for creating brand embeddings (creating embeddings for company names)
0
0
3
useful read: Learning to Learn from Weak Supervision by Full Supervision
Interesting idea for everyone struggling with a lack of labeled data. Has anyone tried something similar successfully before? Day 5
#100DaysOfCode
#DeepLearning
#DataAugmentation
0
2
3
Check out this Meetup with LA Machine Learning!
Hope to see you guys there...
#NLP
#ML
@lamachinelearn
0
2
2
useful read: Document Similarity for Texts of Varying Lengths via Hidden Topics …
Interesting approach: reconstructing wrds in a doc & summary using hidden topic vectors and comparing the approach to Doc2Vec & WMD.
#DeepLearning
#NLProc
#100DaysOfMLCode
0
1
3
We evaluate IIW model-generated descriptions for compositional reasoning. IIW is able to improve accuracy on vision-language compositional reasoning benchmarks ARO and Winoground by several points compared to prior work by generating finer grained descriptions. 🧵9/12
1
0
3
@gensim_py
making life easier for doing
#deeplearning
. jst trained
#wordembeddings
for
#japanese
on
#wikipedia
data. super easy! thanks!
0
2
2
attending
@NipsConference
next week... let me know if you want to meet up. Looking forward to it.
#NIPS2017
#NLProc
#AI
0
0
2
Block 185... Looking gorgeous in the sunset... Google Tower... Austin's Tech Expansion in it's full pride!
#atx
#googletower
#block185
#downtownatx
#google
#googleoffice
#sunset
#austintx
#austin
#austintexas
#tx
#texas
@google
#pixel6pro
#pixelphotography
#googler
0
0
2
any examples of learning material on the working and application of adversarial networks and/or reinforcement learning on text based data?
#GAN
#ReinforcementLearning
#DeepLearning
#NLProc
@sirajraval
0
1
2
@jasonbaldridge
Apologies to Tiger!! Tiger we did not forget you.
image index 88 has the rest of the crew
0
1
2
business model or a winning model.time to mke a decision
@FenwaySportsMgt
for
@LFC
..wastng mney on transfr to send thm on loan..isnt wrking
1
1
2
@vishay
@arankomatsuzaki
Haha we honestly did consider that for a while 😜
There is ongoing work... Stay tuned!
0
0
2
well put in a simple way...
How Natural Language Processing Will Change the Way You Work
#NLP
@salesforceiq
0
2
1
📢 Excited to unveil our latest research, ImageInWords (IIW)! 🚀We're pushing the boundaries of image descriptions with a new seeded, sequential, human-in-the-loop approach producing SOTA, articulate, hyper-detailed descriptions.
arXiv: 🧵1/12
5
32
127
0
0
1
recommendations for a
#NeuralNetworks
architecture for document level classification?
#nlproc
#DeepLearning
#NLP
#ML
0
0
2
“Train Your Machine Learning Models on Google’s GPUs for Free — Forever” by
@nick_bourdakos
…
#DeepLearning
#GPUs
@Google
#Colab
#MachineLearning
#AI
0
0
2
@shao__meng
@GoogleResearch
@GoogleDeepMind
@UW
@AndreaDBurns
@YonatanBitton
@yasumasa_onoe
@RanjayKrishna
@jasonbaldridge
thank you
@shao__meng
for sharing our work! Feedback welcomed.
for reference, full 🧵:
📢 Excited to unveil our latest research, ImageInWords (IIW)! 🚀We're pushing the boundaries of image descriptions with a new seeded, sequential, human-in-the-loop approach producing SOTA, articulate, hyper-detailed descriptions.
arXiv: 🧵1/12
5
32
127
1
0
1
@fly51fly
thank you
@fly51fly
for sharing our work!!
We aim for the work to push the boundaries forward towards mode holistic image descriptions. Feedback welcomed.
for reference, full 🧵:
📢 Excited to unveil our latest research, ImageInWords (IIW)! 🚀We're pushing the boundaries of image descriptions with a new seeded, sequential, human-in-the-loop approach producing SOTA, articulate, hyper-detailed descriptions.
arXiv: 🧵1/12
5
32
127
0
0
0
thank you
#BrendanRodgers
for your services...u almost made it happen. Buts its time for a change...time for
@LFC
to regain its pedigree.
0
0
1
@NBFootball
@LFC
@LFC
u need a microscope to find the difference from previous couple of years...disappointed...look at kits frm other teams
0
0
1
trying out the `.Dataset` functionality for text data. any pointers to a complete code snippet for a network with variable length inputs to be fed into a placeholder?
#AskTensorFlow
@TensorFlow
@mrry
0
1
0
wht a match! 2-2 it is!
@luis16suarez
the culprit and then the savior... the incident looks bad though...
#lfc
#livche
@LFC
0
0
1
Get started with TensorFlow high-level APIs by
@random_forests
@lmoroney
...great talk and really helpful.
#TensorFlow
#io18
#googleio2018
@TensorFlow
0
1
1
another legend retires..u will always be missed...
#thankscatta
#lfc
@LFC
@LFCIndia
@DelhiKop
@LFCMumbai
0
0
1
1
0
1
#Sturridge
&
#Suarez
!!!!
@LFC
amazing counter attacking football !!
#LFC
#YNWA
@DelhiKop
@LFCUSA
@LFCIndia
@LFCMumbai
0
0
1
0
0
1
I never thought I would say this but
@SouthwestAir
you have disappointed beyond repair.. flight getting cancelled with the crew onboard, passages queued up for boarding and that too on
#Christmas
1
0
1
@SouthwestAir
are you expecting passengers to stand in lines for hours to rebook flights with no clarification or update on what to even expect(a hotel credit/voucher?Flight in 2 days?) ? Since online rebooking doesn't work on ur website or app currently and customer care is MIA!
1
0
1
@RealJosephus
maintaining quality across different image distributions is important, avg human struggles as well across domains they are less familiar with.
you bring up a completely valid point.
part of the purpose of the paper was to get this conversation started 👍
0
0
1
memorable day...
#USC
beat
#Stanford
... close game but the underdogs cause the upset...
#GoTrojans
#FightOn
#USC
#USCvsSTAN
#USCFOOTBALL
0
0
1
another week... another win ...
@LFC
top of the league....
#lfc
#liverpoolfc
#epl
@LFCUSA
@LFCIndia
@DelhiKop
@LFCMumbai
@losliverpool
0
1
1
0
0
1
@RealJosephus
absolutely!
@RealJosephus
thank you for exploring and looking forward to how you use it in the future!
length sadly is a very fickle metric, one can have a lot of filter phrases like `this image shows`, `in this image we see`, etc which add no value/richness but inflate the
0
0
1
♫ Pursuit Of Happiness - Extended Steve Aoki Remix (Explicit) – Kid Cudi,MGMT,Ratatat
http://t.co/8FjEx3xExg
#NowPlaying
0
0
0
0
0
1
13 more days...!!! Aug 17th waiting for u... EPL 2013-14..
#epl
#lfc
@LFC
@DelhiKop
@LFCMumbai
@losliverpool
0
0
1
0
0
1
Customer care is down... Line to rebook is 4hr+... I stood in it for 2hr, left in disgust..came back for curiosity after 2hrs of rampant attempts to book another airline and still saw the folks who stood in front me still in line. I have DMed you our flight details
@SouthwestAir
1
0
1
@mikeyoung44
@aimodelsfyi
thank you
@mikeyoung44
! There are some inaccuracies that should be fixed to avoid confusion🙂
Sizes of the released datasets:
For coverage bias: the datasets are indeed collected from diverse domains/topics.
Use-cases: refer Section 4.5 & 4.6
1
0
1
0
0
1
this is exciting, looking forward to all the courses being released!
#deeplearning
@coursera
@AndrewYNg
0
0
1