Tomorrow (10:45am GMT +2) I am presenting Neuroformer at ICLR (#68). Stop by to hear how we trained a multimodal GPT on data from a mouse playing a VR game! 🧠🤖

With all of the excitement of the past few months, it's time for a career update: 🎉I graduated with my PhD from UCSB NLP Group UC Santa Barbara 🥳and joined SynthLabs 🎊to drive open-science collaborations and push the boundaries on data strategies for synthetic data

👇I'm at #ICLR !

UCSB NLP (and alums) at #MASCSLL

Just had an amazing visit to the legendary Alice Oh at KAIST! Blown away by the brilliant students and world-class faculty. It truly is the MIT of Korea! 🌟🎓 #KAIST

had a lot of fun presenting Neuroformer at Caltech this week, thanks for the engaging discussion! 😁

🚨We've been assessing T2I metrics wrong...until now‼️

Our new meta-metric for T2I faithfulness metrics, T2IScoreScore (TS2), checks if a metric correctly orders and separates many images against single prompts!

The results may surprise you...

t2iscorescore.github.io

1/5

Happy to announce my upcoming South Korea 🇰🇷 tour next week 🤩

KAIST - Monday 4/15, 2:30pm E3-1, Rm 4443. Host: Alice Oh
SKKU - Suwon, Tuesday 4/16, TBD, Engineering Hall 2. Host: JinYeong Bak
SNU - Wednesday 4/17, 1pm see below. Host: Jay-Yoon Lee

I hope to meet new + old friends!

When LLMs make mistakes, can we build a model to pinpoint error, indicate its severity and error type? Can we incorporate this fine-grained info to improve LLM? We introduce LLMRefine [NAACL 2024], a simulated annealing method to revise LLM output at inference. Google AI UCSB NLP Group

Huge congratulations to Alon Albalak for defending his PhD thesis “Understanding and Improving Models Through a Data-Centric Lens”. It’s refreshing to witness Alon’s growth, innovation, and leadership in the last few years. Alon is my 8th PhD graduate and I wish him all the best!

🤩 I'm honored that insights from our data selection survey are being shared across the globe 🌍

Fantastic slides Thomas Wolf !

Awesome new work with Sharon Levy and colleagues. I learned a lot about how to think about bias in LLMs from this study.

Niloofar (Fatemeh) @ICLR 🇦🇹 visits us and SoCal #nlproc peeps and shares her work on privacy considerations for large language models.

I'll be uclanlp tmw & UCSB NLP Group the next day to talk about how the 'emergent' capabilities of LLMs create emergent inference-time privacy risks, and how membership inference attacks can be inconclusive in current setups! Hit me up if you wanna chat!

tinyurl.com/mia-cnfide

🚀Fire linguists in the LLM era? No! Excited to share LingoLLM, a novel method for processing endangered languages with linguistic resources. LingoLLM improves the translation of many endangered languages from 0 to 10.5 BLEU! It helps other tasks as well!
arxiv.org/abs/2402.18025

{UCSB|AI2|UW|Stanford|MIT|UofT|Vector|Contextual AI} present a survey on🔎Data Selection for LLMs🔍

Training data is a closely guarded secret in industry🤫with this work we narrow the knowledge gap, advocating for open, responsible, collaborative progress
arxiv.org/abs/2402.16827

[New paper!] Can LLMs truly evaluate their own output? Can self-refine/self-reward improve LLMs? Our study reveals that LLMs exhibit biases towards their output. This self-bias gets amplified during self-refine/self-reward, leading to a negative impact on performance. UCSB NLP Group

Thank you for joining us yesterday Yanai Elazar @ICLR! We are very excited by the 'What's In My Big Data' direction! wimbd.apps.allenai.org

Happy to share our new preprint on understanding how reasoning emerges from language model pre-training: arxiv.org/abs/2402.03268
We hypothesize that language models can aggregate reasoning paths seen in pre-training data to draw new conclusions at inference time.

Congrats to our very own Matthew Ho (Matt Ho) for winning a Computing Research Association Outstanding Undergraduate Research Honorable Mention!
cra.org/crn/2024/01/ou…

💻 LMs leveraging generative pretraining learn many diverse skills.

🧠 But what can GPTs trained on brain data learn to do?

Introducing Neuroformer. A generative model pretrained on massively multimodal, multitask neuronal data! (ICLR 2024) 🧵