Sign In

Paolo Papotti

Associate Professor at EURECOM

Paolo Papotti is an accomplished researcher with a focus on Databases and Data Management.

His current research interests lie in information quality, encompassing data integration, data cleaning, and computational fact-checking.

With a background in Informatic Engineering and a Ph.D. in Computer Science from Università degli studi Roma TRE, Paolo has a strong academic foundation.

He has held various prestigious positions in academia and research institutions, including roles at EURECOM, Arizona State University, Qatar Computing Research Institute, and Università degli Studi Roma Tre.

His experience spans from being an Associate Professor to a Senior Scientist and a Scientific Director in different organizations.

Paolo Papotti's research contributions also extend to the design and implementation of innovative systems for web search, data translation across heterogeneous data models, and metadata management.

He has previous experience as a Software Engineer at Consel and as a Research Intern at IBM Almaden Research Center.

Paolo's diverse academic and professional background has equipped him with a wealth of knowledge and skills in the field of Databases and Data Management.

Highlights

Aug 22 · linkedin

Forget monster GPUs requirements on long context scenarios! FINCH compresses LLMs' context by up to 93x, allowing users to fit a book into a LLM with modest hardware requirements! 🚀✨

FINCH is a novel method designed to compress input context by selectively storing only the most pertinent Key-Value pairs in the cache. This innovation allows decoder-based LLMs to handle longer texts more efficiently without the need for fine-tuning.

How does it work? 🛠️ By leveraging the pre-trained weights of the self-attention mechanism, FINCH identifies the most relevant Key (K) and Value (V) pairs, significantly reducing the memory footprint while maintaining output quality.

Performance Highlights 📈 FINCH delivers compression ratios of up to 93x, ensuring comparable generation quality while being faster and more memory-efficient than traditional methods.

✨ Key Results: On benchmarks like SQuAD v2 and LongBench, FINCH consistently excels in tasks such as question answering, summarization, and code completion, demonstrating high accuracy and faster execution times.

Plug-and-Play Solution 🔌 One of FINCH's features is its seamless integration with any decoder-based model. There's no need for retraining or fine-tuning—just specify your target compression size, and FINCH dynamically optimizes for you. This makes it easy to tap into more efficient LLMs without the usual hassles.

Why It Matters 📊 Large Language Models, like ChatGPT, often face GPU memory constraints and they are limited by the context window defined during their pretraining. FINCH addresses this challenges by optimizing resource usage and let the long context fit within the model context’s window.

Paper: https://lnkd.in/d4XKtTQ7
Code repository on GitHub: https://lnkd.in/d5eM4r7g

Our compression method will be featured in Transactions on ACL (#TACL) and showcased at #EMNLP2024. This work marks the first milestone in Giulio Corallo's Ph.D. journey at SAP under the CIFRE thesis program. 📚🔬

#AI #NLP #Transformers

P
Add to my network

Location

Antibes, Provence-Alpes-Côte d'Azur, France