Dr. Aditya Raj’s Post

Cornell University: "Scalable MatMul-free Language Modeling" by Rui-Jie Zhu and team. They found a way to eliminate matrix multiplication in large language models while keeping performance high. Their method reduces memory use significantly and even outperforms traditional models at large scales. They also created a GPU-efficient version that cuts memory use by over 10x. This work is pushing LLMs closer to brain-like efficiency. #AI #MachineLearning #LanguageModeling #Efficiency #Innovation #CornellUniversity https://lnkd.in/gJcTHWQy

I just stumbled across this promising paper on multiplication free transformer models, and it got me thinking about how we can rethink the way large language models are built. So, what if we could create high-performing AI models without relying on the computationally expensive matrix multiplications that currently dominate the process? Here’s the breakdown: • Matrix multiplication (MatMul) is a core operation in almost all neural networks, but it’s extremely costly in terms of memory and processing power, especially when scaling up models with billions of parameters • The authors of this paper propose a novel approach: replacing MatMul operations with ternary weights (values constrained to -1, 0, and +1). Instead of complex multiplications, we’re simplifying things down to basic additions and subtractions. They’ve managed to significantly reduce memory usage—up to 61% less during training and 10x less during inference—while maintaining competitive performance with state-of-the-art models. 🤯 Who would have thought ... could this be the future of edge AI? https://lnkd.in/eBszQWwY

More Tech for Good Research - Nature PrePrint Chemistry specific Agents fuelling domain specific capability whilst leveraging the power of LLMs. Large language models can be queried to perform chain-of-thought reasoning on text descriptions of data or computational tools, which can enable flexible and au… #TechforGood - reasons to be cheerful. #AI Source: Nature

SCALABLE MATMUL-FREE LANGUAGE MODELING Failure to pay attention to this paper could mean missing out on a breakthrough that could shape the future of AI. This paper, which addresses the replacement of matrix multiplication (MatMul) operations in large-scale language models (LLMs) with matrix addition (MatAdd) operations, should not go unnoticed, due to the enormous potential impact it can have on the development of AI, both at the theoretical level (R&D&I) and at the level of product development. This approach has the potential to outperform current methods by a significant margin, paving the way for a new generation of AI models. Just a data point: With this new method, efficient GPU implementation would reduce memory usage by up to 61% and accelerate LLM training by 25.6%.

I am excited to share my latest article on the Transformer architecture for Large Language Models (LLMs)! If you are curious about the technology behind advanced language models, this article provides a comprehensive overview of the architecture and its applications. #ArtificialIntelligence #MachineLearning #NaturalLanguageProcessing #Transformers #DeepLearning #AI #DataScience #LLMs #TechInnovation #AIResearch

"Innovation starts with mastering the fundamentals." Specifically talking about Generative AI, the intricate workings of large language models, their mathematical foundations, and the techniques that power them form the backbone of every breakthrough in the field. Without a deep understanding of these core principles, it’s hard to push the boundaries or optimize existing systems. True innovation comes from grasping the "why" behind the "how." With the increasing reliance on directly calling APIs, it’s easy to lose touch with the core principles that make large language models work. While APIs provide quick and convenient solutions, they can limit our ability to think critically and innovate beyond what's already available. I recently came across an exceptional resource that brings these concepts together in one place. It explores the mathematical workings of LLMs, training methodologies like pre-training, fine-tuning, and alignment(RLHF/DPO), training optimizations, context length extension techniques, prompt compression techniques and many more. For anyone working on generative AI, this book is a must-read. https://lnkd.in/gTTwdWyT

Scalable MatMul-free Language Modeling is a potential breakthrough in how AI is performed, aiming to make AI less brute force. For those interested in delving into technical details, check out the work at https://lnkd.in/gkGkJvaQ.

Amid the excitement of two Nobel prizes awarded to AI leaders this year, my first GPT paper is out! (No, we did not write it with GPT... or did we? 🤔 Who can tell these days?) All jokes aside, my colleagues Xin Wang, Liangliang Huang, Kun Lu, and I explored the ability of GPT-4 to extract band gap energies from 415 abstracts. With the right prompt engineering, GPT-4 achieved an impressive 94% accuracy, a huge leap from the 51% accuracy we saw using a traditional rule-based method. Check it out and let us know what you think! #AI #GPT4 #MachineLearning #InformationExtraction #BandGap

Updates on #AI : Unlocking Intelligence: Solomonic Induction for Large Language Models, Full article link 👇🏻👇🏻 https://lnkd.in/drbWY9A7 #artificialintelligence #machinelearning #ML

Just finished the course “Introduction to Prompt Engineering for Generative AI (2023)”!  - Prompt engineering? - Tokens v. words - Large Language Models - Text Generation - AI-Generated Images - Fine-tuning your prompts - APIs Check it out: https://lnkd.in/e_KbYJj2