AerIn
@aerinykim
Followers
6K
Following
17K
Media
145
Statuses
4K
building https://t.co/rjIQoeDYwX. honoring minds built on irreducible foundations https://t.co/JVijyuw4Z2
Joined November 2014
we've been supporting 's initiative to identify 2024 election deepfakes for a year. here's a peek at some interesting submissions we've had. when in doubt, check on . we're hopeful for this Tuesday's outcome!.
2
0
8
Are you a PhD student or an independent researcher who is working on creating a dataset?. @scale_ai will sponsor promising dataset creation projects that can widely benefit a community. Please DM me for details.
21
205
827
So far (2 days gone), this is my favorite #NeurIPS18 paper. I just love the paper that tells me why it works. It also showed that the popular previous belief "internal covariate shift" has little to do with batch norm's success. Ah! 😍. 3-min video:
5
213
783
Training Data. The neglected in academia, but an important topic in industry. I learned the hard way- errors in training data can easily wipe out accuracy gain from a model. I'm writing codes to monitor data fed to ML. Things to check (can you add more?):.1. skewness between.
15
131
643
I still remember being surprised 8 yrs ago when I first reduced Word2vec/Glove embeddings to a lower dimensional space. To my surprise, antonyms were placed very close to the words themselves and models sometimes failed to understand negation. At the time, I couldn't find a way.
You might have heard that word2vec, test of time award winner at #NeurIPS2023, was rejected from #ICLR2013 back in the day. Interestingly, one reviewer felt so strongly that they recommended "Strong Reject" four times.
10
47
567
My team spoke very highly about this blog (and they're also wondering if self-supervised learning could eliminate the need for labeling entirely) so I gave it a read. It was a very well-written, thorough overview of SSL. What stands out the most was it was written by Dr. LeCun,.
5
50
460
What @bryan_johnson does is pretty common for South Korean women. As a woman who was born and raised in Korea, I've been doing basically the same, like religiously avoiding the sun, regularly getting facials, calorie restriction, etc. It's the norm there.
26
16
453
My new boss 🕶️🚬 got me a new desktop from @LambdaAPI. 4x @Nvidia RTX 8000 (48G VRAM in a single GPU. Total 192 GB).I was able to wrap them in a beautiful box. I certainly cannot wrap 8 Titans in a box. Which model should I try first on my new machine? ☺️
30
27
380
Before I forget, I'd like to summarize some interesting papers that I found at #CVPR2022. Dual-key multimodal backdoors for visual question answering. 1. This paper proposes an interesting Trojan attack method. To start, what exactly is a Trojan attack?
6
50
296
One thing that stood out to me at #NeurIPS2023 was there were a lot more application papers than before. This was not the case with my last neurips (2018). I'd like to encourage this wave by highlighting one such paper. LayoutGPT: Compositional Visual Planning and Generation
2
38
249
How can we tame GPTs? .Tame as, how do we make LLM give us more of what we asked for, instead of random answers?. Also, how do we discover what GPTs are capable of? .Can we use LLM as, e.g. copilot or translator?. We'll discuss I-GPT, which all OpenAI GPT production APIs use.
How can you ensure your AI language models are aligning with human intent? 🤖. Join Long, OpenAI Research Scientist, and Aerin, Scale AI Engineering Manager, for a technical deepdive followed by discussion around Long's work with InstructGPT, and live audience Q&A.
0
0
8
"Engineers over 35 aren't sharp/fast anymore.".-> Not true. Attended @Replit hackathon last week and teamed with a former sr dir of eng at xxx, 35+. He was one of the fastest and most effective I've ever worked with. I don't think coding skill is something that fades with age.
12
14
208
Introducing a new paper: Natural Adversarial Object. published @AndrewYNg's Data-Centric AI workshop in #NeurIPS2021. What is a Natural Adversarial Object?🤷. You might have already heard of adversarial samples - the inputs to ML models that we intentionally created to confuse.
🦈Natural Adversarial Objects: a new dataset to evaluate the robustness of object detection models. NAO contains 7,934 images and 9,943 objects that are unmodified and representative of real-world scenarios, but cause SoTA detection models to misclassify.
3
48
196
I wrote about the Beta distribution during the holidays. 1. Why does the PDF of Beta dist look the way it does?.2. Why do we want to use Beta dist?.3. Intuition for its shapes . Feedback welcome! I'm still looking for an explanation of U-shaped Beta.
7
22
186
When linear bmm made LSTM disappear, I thought some other overly elaborated architectures will fade away too. But I didn’t expect CNN belongs to that category. .
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale . When pre-trained and transferred to CV tasks, Vision Transformer attains excellent results compared to SOTA CNNs while requiring much fewer computational resources to train.
4
19
164
Why can Lagrange multipliers for inequality constraints only be positive? . A blog post (more like an article) beautifully visualized using @pyray1 -- a math visualization tool I'm also contributing to!.
0
24
142
Having long-time unanswered questions about Poisson distribution?. e.g. -Why does the formula look the way it looks?.-Why did the Poisson dist have to be invented, even though we have binomial?.-or just confused with λ and k?. then this blog is for you!.
2
6
98
Hi Twitter around the globe👋. Please join my talk at 1pm PST, Wed 6 Oct @Scale_AI TransformX!. right before @hardmaru's talk. I'm going to talk about ML models we built using LIDAR, our recent @NeurIPSConf dataset submission, SOTA annotation tools, etc.
0
18
103
We often frown upon the concept of manual feature engineering,. but if you look at *any* NLP model (BERT, transformer, etc., you name it!) that we are using now,. their network structures are highly manually engineered too! ;).
2
6
93
Almost all COVID-19 clinical studies use Chi-square test and P-value. ". when the coronavirus patient shows signs of dizziness (p-value 0.007), should he be sent to ICU, our most important resource?". Here is how to correctly interpret them.
3
23
88
@abbyvesoulis @abbyabrams In South Korea, the cost of 19 days of hospitalization for complete corona treatment is $9700. Every single patient's out of pocket cost is $44. The government pays the rest.
1
24
81
@srush_nlp @stanfordnlp @huggingface @Thom_Wolf Thank you! Huggingface is doing what openai was supposed to do.
1
2
59
training vs serving data (imo this affects # the most).2. outliers in training data.3. data drift: how data changes over time.4. fairness: sample size disparity, unwarranted association, proxies.5. robustness: some changes (noise, perturbation, etc.) to the input should not.
3
2
56
After attending ICLR/NIPS, I conclude Research (or tech in general) compared to other industries, is a place of true meritocracy, transparency, accountability and integrity. People who possess those qualities flourish here. I'm constantly awed and inspired by humans in this field.
0
4
55
I wrote a short article on Operations Research, one of the most fastcinating and amusing subjects in applied math (but not that well known major in universities) and also one of the most widely used techniques in Machine Learning!.
Operations Research — what, when and how by @aerinykim
3
3
51
change classifier’s output.6. case level testing .7. differential privacy: whether any one individual's data has a significant effect on outcome.8. data versioning: auditability and traceability.9. data / model relevance: don't overcomplex the model than it's necessary.
5
2
48
My founding team (the first 4) will be all women. I think it's possible.
6
0
46
I'm so happy that the holidays are coming because I can study. Am I the only one who feels this way?.
6
0
47
So far the best #ReinforcementLearning blog post that I've read in 2018. (Disclaimer: I put my two cents here.) So clear and so original.
Decided to write up something a little different from my usual posts: Introduction to Learning to Trade with Reinforcement Learning.
1
7
45
Wrote a blog detailing what made RLHF/ChatGPT so powerful and why calling it emergent AGI is a slight overstatement. 1. What exactly has changed between ChatGPT and Transformer.2. What made it too helpful (hallucinate).3. Action items for Open-Source LLMs.
1
7
47
We need to treat data more seriously in ML pipeline, on par with the architecture and framework. It will affect the applicability and scale of your research. A now-deleted anonymous tweet says
1
3
45
@geoffreywoo @bryan_johnson exactly. i don't know about trt but peptides and ipl have been used for at least 20 yrs. IPL is now out of fashion and Korean women started using polynucleotide. we can't get that here because the FDA hasn't approved it.
0
0
38
The best AI products can't be built from boxed solutions. Your engineers would need to customize the data/models (not through YAMLs!) and they need very transparent tools, not just cloud services.
2
6
36
As some of you already know, today is my last day at Microsoft. (also the day that President Trump tested positive). I’m very proud of what my team has accomplished. We built a world-class QA product with the most powerful set of AI capabilities on enterprise structured database!.
1
0
35
Hi Twitter👋.Come to my tech talk tmrw @scale_AI exchange!. I'll discuss our new stance detection dataset (Reddit 50k) and how to make the best use of it. Spoiler: predicting stance can be a graph problem.
2
3
33
Went to a friend's birthday party, met 3 founders, and found out we are all now working on search. (because openai solved the conversation layer.). It's interesting to see search is being talked about again and there still isn't a good search API. (is it because it's too valuable.
4
1
35
Another exemplar, informative talk by @goodfellow_ian “Defense against the Dark Arts” #CVPR18
0
7
33
Our Reddit 43K: A comment-reply dataset for dis/agreement detection paper just got accepted inaugural @NeurIPSConf dataset track!. from amazing & dedicated @lejohnyjohn6, Valentina Semenova, @__alexmatton, Rachel Han, @RenaudLambiotte, Doyne Farmer. paper:
1
11
32
3
3
31
#CS224n is such a useful mooc for NLP startups! Probably the best #deeplearning application mooc available.
Excellent turnout as @RichardSocher kicks off #CS224n: Natural Language Processing with Deep Learning at #Stanford!.
1
5
27
Some personal news:. It’s been an amazing journey helping Scale reach its goals over the last two years, but it is time for me to pursue something different. During my time at Scale, we accomplished some remarkable things, such as building the largest revenue product that didn't.
1
0
29
Bidding on Neurips dataset work, I'm so excited to see real-world data from organic needs that is closely tied to the applications. Really appreciate authors' decision to make the data accessible to a larger audience. Some highlights:.- Skin disease data (densely annotated).
4
4
26
I know I'm going to lose 100+ followers by retweeting this. but this number needs to be improved. I know many capable women (not only in tech) constantly getting low-value work, ended up not having enough time to develop oneself. People who spend more time on low value tasks are.
The [gender] gap in artificial intelligence is three times larger than in other industries . Women with AI skills are more likely to be employed as data analysts and information managers, while men tend to land in more lucrative and senior positions.
4
5
24
Whenever I find a person to be exceptionally mature, or thoughtful, or kind, almost always, their family member is ill or has some kind of special needs. In that sense, maybe living with some form of hardship is what makes a better human being. maybe.
2
1
21
I'll be presenting a RLHF session at AI Summit Seoul on the 14th, starting at 2pm KST. We'll be diving into the intricacies of instruction data prep, so come hang out!.
3
2
23
Having a very pleasant virtual experience at #ICLR2020!.Zooming to discussions actually works 👍. The functionality that I like the most is the *similar papers* on each paper. (why didn't we have this before?). Thank you for the amazing work! @iclr_conf.
0
2
22
After tax day thoughts:. When 30% to 50% of ppl's paychecks are taken for taxes, shouldn't we expect some basic safety in return? 🤔. The US gov isn't even responsible for healthcare. So I'm interested what's keeping the SF local gov from providing some level of safety.
3
1
19
Why a human-in-the-loop auditing tool is necessary. Good read:
We show this in our work on compression. Pruning and quantizing deep neural networks amplifies algorithmic bias. and
0
6
21
This is why @wimlds praise @goodfellow_ian so much. A genuine person who helps others. Fun Facts: He applied for an internship and Stanford downgraded his transcript. In mathematics, none of the four most recent Fields Medal winners has over 5,000 citations.
0
3
20
1
0
21
A great way to kick off #NeurIPS2020! . Check out the @scale_ai EXPO talks: . Dec 6, 4pm PST ."Scaling Data Labeling with ML" by @phelixlau. 6pm PST .“Visually Debugging ML Models with Nucleus” by Elliot Branson. We will be around for questions after the talk. See you there!.
0
2
18
Come visit our poster #221b at #CVPR2022 today!.Recognition: Detection, Categorization, Retrieval. See how we handle multi-category attribute predictions (e.g. color, moving/not moving, signal flashing) in detected objects (e.g. a vehicle)!. @go_krm.@elliot_branson .@scale_AI
0
4
18
I'm giving away free #AIWTB tickets! Retweet!.@goodfellow_ian, Sam Altman, Yoshua Bengio, Siraj and I are speaking
3
16
19
"GlideNet" accepted at #CVPR2022. I'd like to highlight some interesting aspects of our work. 1. In real-world applications, it's critical to assign visual attributes (color, shape, motion, etc.) to detected objects. This is usually formulated as a multi-label classification
2
3
16
Where do you get training data for your ML applications?. 1. From research dataset. (then diligently tweaking data).2. Collect from our app (I guess you already have sizeable app up and running?).3. Buy from 3rd party (from who? Upwork? Pls comment).4. Hire Mturks and author w em.
4
4
15
While I don't personally gain anything from OpenAI's success, watching their success unfold over the years is uplifting!. To me, it's a clear demonstration of the power of meritocracy. People might think OpenAI only hires those with highest credentials, but that's not true.
2
1
17
@aleonardobecerr @tkasasagi This is what to look for in your future partner (listed in order of importance). 1. kind & generous.2. healthy.3. financially responsible .4. can have sex.5. love children.
4
0
17
"Is word2vec supervised or unsupervised? It is supervised in a sense that we punish wrong predictions during the training but the corpus is actually not labeled. So it's unsupervised as well.". No one in the room answered my question saying "it's called self-supervised learning!".
3
1
15
I'd like to share the work of one of the best scientists that I know!. @DreamwaveAI. I uploaded only 5 really boring, no-makeup images, and it gave me 200+ model shots with different clothes and backgrounds. Reply👇 for an April Fool's discount coupon!.
0
2
15
They made every detail available to the public. It was much-needed best practice. Love this work.
Today Meta AI is sharing OPT-175B, the first 175-billion-parameter language model to be made available to the broader AI research community. OPT-175B can generate creative text on a vast range of topics. Learn more & request access:
0
4
16
This is why you'd want to recruit an ml engineer with taste.
i try not to think about competitors too much, but i cannot stop thinking about the aesthetic difference between midjourney and dalle
2
0
16
Now, who is delivering the world's most groundbreaking AI products?. A team of hungry doers .in an org with relatively less bureaucracy!. I've always suspected this to be true, but I hadn't seen a clear real-life example until now. I want to build a team like that.
2
0
16
This is why we should always choose open over closed. Look at what has happened in the last two weeks since the StableDiffusion team decided to release the model in its entirety.
Surprised to be able to run #StableDiffusion locally on my M2's GPU in under 30 minutes. Definitely feels like a kind of 🚀 launch to me — I now get it.
0
1
15
exist before and run many successful pilots. I'm looking forward to seeing Scale to further grow. As for me, I want to try out a few of my ideas and spend more time on my passion around education. I'll be in SF until this month, DM me if we should meet or if you work in Edu!.
2
0
15
This is a Youtube channel that I watch these days when I eat alone: ."TechLead" by @PatrickShyu. This guy is hilarious. And profoundly insightful too!. "IRL, there is no binary tree to traverse".
1
3
12
saw this poster and thought of @tkasasagi. Document Enhancement using visibility detection #CVPR18
1
4
15
1
0
14
Watching Tokyo Olympics this summer, especially @AnnaKiesenhofer, the Austrian Mathematician who won the gold medal in cycling, made me think the concept of training is not unique to pro-athletes but also applies well to other technicians, e.g. engineers, musicians, dancers, etc.
1
0
13
Why make a dramatic prediction in such a short time frame?.In my experience, working without drama leads to better outcomes every time.
By 2024 you’ll be able to replace ~50% software devs with GPT-4 agents that run on $10 worth of tokens per hour. The whole “they don’t need sleep or breaks or food” thing?. Yeah. That’s real now. Why hire a new employee when you can spin up an AI agent for 1/10 the cost?.
5
0
14
Love this software engineer's explosive participation in ML since the chatgpt era. Before chatgpt, swes weren't very involved in ML or in developing AI products. Now it's swe's time to drive AI product advancements!. LOVE their velocity and enthusiasm, and this is healthy.
2
0
14