Ritesh Sarkhel

Stuff that I was working on for the past year but couldnt talk about publicly. Hello Rufus 🐶 nytimes.com/2024/02/01/tec…

Rufus 🐶 is here to help you on Amazon if you're shopping from North America today. Better attribute search, open-ended QA, product recommendation, and general chitchat all in one app 💫 aboutamazon.com/news/retail/ho… #LLM #Amazon #Rufus

hi rufus ...

@keviv9 @SCAI_ASU @ASU @CityofPhoenixAZ Great news! Congratulations 🎊

I’m beyond excited to share some amazing news: I've accepted an Assistant Professor position at @SCAI_ASU Arizona State University @ASU in Tempe, AZ! 🌵🎓 I'll be starting this thrilling new chapter from Fall 2024. Phoenix @CityofPhoenixAZ, here I come!🚀🌞(Mr ->Dr. ->Prof.) -1/3

@yunyao_li Happy to help on either task. Please feel free to DM if you're still looking for emergency reviewers.

Dear all, I need a few emergency reviewers for manuscripts related to (1) handwriting recognition; (2) common NLP tasks (e.g. QA, NER). If you have strong publications in these areas and have bandwidth to review within the next 2-3 weeks, please DM me. Thanks.

@infoxiao I was going to make the same joke haha

Okay one more layer -- Announcing NeurIPS Moms track!

Wow!

you: i use Claude 3 Opus for coding me: i use the Amazon Shopping app for coding

[CL] Noise-Aware Training of Layout-Aware Language Models R Sarkhel, X Ren, L B Costa, G Su, V Perot, Y Xie, E Koukoumidis, A Nandi [Google & The Ohio State University] (2024) arxiv.org/abs/2404.00488 - The paper proposes a Noise-Aware Training (NAT) method to train layout-aware language models for information extraction from visually rich documents in a scalable way. - NAT utilizes weakly labeled documents supplemented with limited human-labeled documents to train the model, avoiding expensive human annotation effort. - To prevent performance degradation due to noisy weak labels, NAT estimates the confidence of each training sample and incorporates it as an uncertainty measure during training. - Experiments show NAT-trained models outperform transfer learning baselines in terms of macro F1 score while requiring significantly less human labeling effort. - Key aspects of NAT include sample reweighting, weight thresholding, noise-aware loss, and sequential fine-tuning on corpora augmented with weak and synthetic labels.

NAT introduces a systematic way to train layout-aware #LLMs on noisy documents. It reduces labeling cost w.o. drop in perf and works in multi-lingual settings out-of-the-box. Thank you for the shout out @_akhaliq. This was truly a labor of love during my time @GoogleAI

@kdd_news More details about the competition rules, the prize pool, and the dataset is here: aicrowd.com/challenges/ama…

@kdd_news It also gives a peek of 🏋‍ShopBench 🏋, a massive #LLM evaluation benchmark curated in-house to mimic the nuances of real-world online shopping complexities.

📢 We're hosting a @kdd_news Cup competition & giving away cash prizes ✨ The Massively Multi-Task Online Shopping Challenge invites #LLM researchers to try their hands on a set of tasks that has an outsized impact on online shopping experience (1/n) #LLM #GenAI #LLM #AI #Amzn

@pratyushmaini @goyalsachin007 Curious to know your thoughts on the performance of a model trained this way on OOD datasets

An exciting data curation paper came out from Google. I had to call @goyalsachin007 because the results challenged my prior beliefs about web scraped data quite dramatically. Read his thread to see what we think is happening. 👀 Into the age of pre-train like you fine-tune (1/n)

Very cool!

1/n: There are some academic papers that are so brilliantly and so accessibly written and so universal in scope that they transcend disciplines and stand as timeless testaments to both great thinking and great writing. Here's a short personal selection:

We're hiring PhD interns and Scientists to work on exciting projects related to #LLMs and #GenAI @amazon. Feel free to reach out if you're interested.

More detailed write up here: aboutamazon.com/news/retail/am…

Stuff that I was working on for the past year but couldnt talk about publicly. Hello Rufus 🐶 nytimes.com/2024/02/01/tec…

Friends don't let friends make bad charts! Chenxin Li, pulled together a lot of great advice for data visualization, with clear "do this, not that" examples for each item. Here are a few of my favorites, see the link below for more.