Foundations of Artificial Intelligence Seminar Series

Speaker Schedule and Abstracts

• 🎤Pavlo Molchanov 🏛️NVIDIA Research 📅Thursday, October 10th, 2024 | 12 PM - 1 PM (EDT) 📍B5 Classroom in Boggs | 💻Zoom 📚Talk Title: Efficiency in Large Language Models with Post-Training Compression
  📝Abstract Training large language models (LLMs) for various deployment scales and sizes traditionally involves training each variant from scratch, a process that is highly compute-intensive. In this talk, we explore three key techniques to significantly enhance LLM efficiency: (1) pruning and distillation, (2) Flexible LLM architecture with the Many-In-One concept, and (3) an advanced Parameter Efficient Finetuning Technique. Pruning and distillation can reduce pretraining costs by up to 40x, delivering models that are up to 16% more accurate than those trained from scratch. The Flexible LLM architecture allows the transformation of a single LLM into an infinite number of smaller sub-models, streamlining deployment across various applications. Lastly, we will discuss DoRa, a state-of-the-art parameter-efficient fine-tuning method based on weight decomposition, enabling efficient model fine-tuning with limited data.
  👤Bio Pavlo Molchanov is a Distinguished Research Scientist and Team Manager at NVIDIA Research. Since 2023, he has been leading the Deep Learning Efficiency Team. He obtained a PhD from Tampere University of Technology, Finland, in 2014. During his studies, he received the Nokia Foundation Scholarship, GETA Graduate School grant, Best Paper Award, and Young Researcher Award at EuRAD. Recently, he has focused on efficiency in LLMs and multi-modal models: compression, NAS-like acceleration, novel architectures, and adaptive/conditional inference. His past research has led to several NVIDIA product integrations: hand, body, and facial keypoint estimation and recognition in DriveIX, Broadcast, Omniverse, Maxine; efficient vision backbones in TAO, developed compression techniques in TAO, NVIDIA AV, TRT Model Optimization; and small in-game LLMs.