Description:

In this talk, we will explore the necessity of fine-tuning pre-trained large models for specialized tasks, starting with an introduction to the conventional fine-tuning methods. We'll highlight the limitations of these methods, particularly regarding their computational and storage demands.

To address these challenges, we will shift our focus to Parameter-Efficient Fine-Tuning (PEFT) methods. Specifically, we will delve into an adapter-based PEFT technique called LoRA (Low-Rank Adaptation). We'll discuss the purpose (the Why) and significance (the What) of LoRA, aiming to enhance our understanding of its methodology and rationale, especially in the context of large models.

Next, we will implement (the How) LoRA in a multilayer perceptron (MLP). We'll start by setting the stage for fine-tuning with a toy dataset for a binary classification task. Then, we'll detail the fine-tuning process with LoRA, including adapter insertion, parameter configuration, and the assessment of resulting parameter efficiency. This will underscore LoRA's ability to mitigate computational and storage demands.

We'll also cover practical considerations for sharing and loading models through the Hugging Face Hub, enhancing the utility and accessibility of fine-tuned models. Finally, we will address LoRA's limitations in memory usage during inference and discuss solutions like Quantized-LoRA. We'll round off the discussion with a comprehensive look at managing and optimizing large language models (LLMs) for specific tasks with minimal resource overhead.

Please check out the blogs mentioned in the "Content url" section.

Prerequisites:

Basic Deep learning topics and basic Linear Algebra

Video URL:

https://drive.google.com/file/d/1IBDlcbWWCPBOuREJ-7SNzby0Xos0x7sw/view?usp=sharing

Content URLs:

Blog 1: https://medium.com/inspiredbrilliance/exploring-lora-part-1-the-idea-behind-parameter-efficient-fine-tuning-and-lora-ec469d176c26

Blog 2: https://medium.com/inspiredbrilliance/exploring-lora-part-2-analyzing-lora-through-its-implementation-on-an-mlp-fbc386036f6f

Speaker Info:

Speaker bio 1 - Preethi Srinivasan is a Solution Consultant at Sahaj Software. She has a Masters (by Research) from IIT Mandi. As part of her masters thesis she worked on applications of Deep Learning algorithms to medical imaging problems. She published her works at NIPS-WIML workshop, IEEE CBMS-20 and ACCV-20 conferences. At Sahaj, she has developed prototypes for Video Summarization (in an unsupervised fashion) and Video Captioning, focusing on extracting meaningful information from the video data. She is currently working on building question-answering systems for specific domains based on document analysis, utilizing RAG and/or fine-tuning on LLMs.

Speaker bio 2 - Shruti Dhavalikar is a skilled Data Scientist with over 4 years of experience in the field, currently working as a Solution Consultant at Sahaj Software in Pune, India. With a deep passion for data science and a strong understanding of data analysis, she thrives on designing models that build valuable business insights from data. She has successfully delivered end-to-end product cycles under Agile methodology, showcasing her ability to handle diverse tech stacks and ensure scalable, clean, and robust design. Passionate about her work, she actively contributes to research projects and tries to stay at the forefront of advancements in the field. She has published research papers at international conferences.

Speaker Links:

Speaker 1 links: Linkedin: https://www.linkedin.com/in/preethi-srinivasan-3a221915/

Google Scholar: https://scholar.google.com/citations?user=JMWfeg4AAAAJ&hl=en

Github: https://github.com/s3pi

Youtube: https://www.youtube.com/channel/UCamYscGMk1gAgI7A4XAwd4g

Speaker 2 links: Linkedin: https://www.linkedin.com/in/shruti-dhavalikar-83514615a

Google Scholar: https://scholar.google.com/citations?hl=en&pli=1&user=5029f7QAAAAJ