🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

📢 New #Master #AI course on #Foundation #Models. We consider fundamental algorithmic questions concerning their #multimodal #learning and #reasoning abilities, as well as the #ethical, #legal, and #social aspects of their deployment in society. uvafomo.github.io

Learn more about data governance with the latest AI techniques. 2nd edition of our 1-day Masterclass Modern Data Governance is coming up on June 6th 2024 led by expert prof. dr. Paul Groth. Sign up now: academy.uva.nl/moderndatagove… #UvA_Science #datamanagement #education LAB42 UvA

Our paper "Towards Interactively Improving ML Data Preparation Code via 'Shadow Pipelines'" has been accepted for the DEEM Workshop @ SIGMOD at SIGMOD! 🎉 In this vision paper, we present our initial ideas for my next research project. Joint work with Sebastian and Paul Groth.

🎉Our paper “Towards Efficient Data Wrangling with LLMs using Code Generation” has been accepted at DEEM Workshop @ SIGMOD at SIGMOD! We envision a hybrid approach that utilizes LLM to generate code and apply on per row basis for data wrangling. Work done at MotherDuck w/ Till Döhmen

And we celebrated the 500th promotion chaired by prof. Peter van Emde Boas. What a milestone! . Thanks from Informatics Institute University of Amsterdam! I’m sure many here on Twitter/x had him as chair during their defense.

Life of a director - week 12

🗺 Are you a BSc or MSc student at the University of Amsterdam and have an accepted paper at a top-tier conference (ML or AI related)? Then you might qualify for the ELLIS/ Qualcomm travel grant! 🎉 🔗 Find out more about the grant and how to apply for it: ivi.fnwi.uva.nl/ellis/wp-conte…

We're pleased to welcome Daphne Miedema and Lise Stork to the lab as new assistant professors. They bring strong research lines to the team: - Dr. Miedema works on human data interaction - Dr. Stork works on data engineering for scholarly data UvA Amsterdam Informatics Institute University of Amsterdam

"AI everywhere, but without sky-high energy usage". Dolly Sapra's research focuses on making large AI models available in devices that operate in isolated environments, keeping energy efficiency in mind. LAB42 UvA Informatics Institute University of Amsterdam. Read more about her work: 🔗 ivi.uva.nl/content/news/2…

Introducing our new paper🥳: ApiQ: Finetuning of 2-Bit Quantized Large Language Model🧐 In short: ApiQ can work as a quantization framework, and a memory-efficient finetuning framework, like QLoRA. Paper: arxiv.org/pdf/2402.05147 Code: github.com/BaohaoLiao/ApiQ (1/n)

Strengthen your IT security expertise! Join our 5-part lecture series during which you will get knowledge and insights to enhance IT security in your organization. Starting November 5th 2024. More info: uva.nl/foundations-of… #UvA_Science #ITsecurity #cybersecurity

LTL will present two papers at ACL 2024: 1. The Fine-Tuning Paradox: Boosting Translation Quality Without Sacrificing LLM Abilities by David David Stap 2. How Far Can 100 Samples Go? Unlocking Overall Zero-Shot Multilingual Translation via Tiny Multi-Parallel Data by Di Di Wu

Paper details: 1. The Fine-Tuning Paradox: Boosting Translation Quality Without Sacrificing LLM Abilities arxiv.org/abs/2405.20089

How Far Can 100 Samples Go? Unlocking Overall Zero-Shot Multilingual Translation via Tiny Multi-Parallel Data: arxiv.org/abs/2401.12413

🚨 New paper 🚨 Our multilingual system for the WMT24 general shared task obtain: --- Constrained track: 6 🥇3 🥈 1 🥉 --- Open & Constrained track: 1 🥇2 🥈 2 🥉 A simple and effective pipeline to adapt LLM to multilingual machine translation. paper: arxiv.org/abs/2408.11512

#NeurIPS decision is out

🚨Want to finetune a 70B LLM on a single A100-80GB without any memory-saving tricks? Come to discuss with us at #EMNLP2024 Main 🗓️Tuesday (Nov 12), 16:00-17:30 📍Poster session 04 #LLM #quantization #PEFT

Check out our work on Reward-Guided Speculative Decoding! 🚀 • Use PRM for reward-guided sampling — a mixture distribution • Prove binary weighting is optimal under budget constraints • Saves 4.4× FLOPs in STEM • Outperform speculative decoding 🔥💡 arxiv.org/pdf/2501.19324

Impressed by DeepSeek-R1 and o3? However, they are long-reasoning models, and generate >4k tokens quite often for hard questions. It’s very time-consumed! Here is our solution to speed up their inference! #deepseek #DeepSeekR1 #o1 #o3 #reward