Hey Folks!

This newsletter will discuss some cool AI research papers, AI tools, and AI Startups. Happy learning!

➡️ A Team of UC Berkeley and Stanford Researchers Introduce S-LoRA: An Artificial Intelligence System Designed for the Scalable Serving of Many LoRA Adapters

How can the "pretrain-then-finetune" paradigm in large language model deployment be optimized for efficient serving? This paper introduces S-LoRA, a system designed for a scalable serving of numerous Low-Rank Adaptation (LoRA) adapters. These adapters, derived from a base model, are stored in the main memory, with the relevant ones fetched to the GPU memory for current queries. S-LoRA's unique feature, Unified Paging, manages dynamic adapter weights and varying KV cache tensors within a unified memory pool, minimizing GPU memory fragmentation. It incorporates a novel tensor parallelism strategy and custom CUDA kernels for heterogeneous batching of LoRA computation. S-LoRA significantly outperforms existing libraries, boosting throughput by up to 4 times and increasing the number of served adapters substantially. This system facilitates scalable serving of multiple fine-tuned models, paving the way for extensive, customized fine-tuning services.

➡️  Google DeepMind Researchers Propose a Framework for Classifying the Capabilities and Behavior of Artificial General Intelligence (AGI) Models and their Precursors

How can we effectively classify the capabilities and behaviors of Artificial General Intelligence (AGI) models? This paper proposes a framework, analogous to the levels of autonomous driving, to categorize AGI models based on performance, generality, and autonomy. By analyzing existing AGI definitions, the authors distill six key principles for a practical AGI ontology, emphasizing capabilities over mechanisms, and the distinction between generality and performance. The proposed 'Levels of AGI' are based on the depth (performance) and breadth (generality) of capabilities. The paper reflects on how current systems align with this framework and underscores the need for robust benchmarks to evaluate AGI models. Furthermore, it explores the interplay between AGI levels and deployment considerations, highlighting the importance of responsible Human-AI Interaction paradigms for the safe implementation of advanced AI systems.

➡️  Are You Doing Retrieval-Augmented Generation (RAG) for Biomedicine? Meet MedCPT: A Contrastive Pre-trained Transformer Model for Zero-Shot Biomedical Information Retrieval

How can we improve information retrieval (IR) in biomedicine without extensive query-article annotations? This paper introduces "MedCPT", a first-of-its-kind Contrastively Pre-trained Transformer model designed for zero-shot semantic IR in the biomedical field. To develop MedCPT, a vast dataset of 255 million user click logs from PubMed was utilized. This data facilitated contrastive learning, enabling the training of an integrated retriever and re-ranker system. MedCPT demonstrates superior performance, surpassing various models including those as large as GPT-3-sized cpt-text-XL, across six biomedical IR tasks. Additionally, it excels in generating nuanced biomedical article and sentence representations. This model offers immediate applicability in diverse real-world biomedical IR applications.

➡️ Researchers at Stanford Introduce CORNN: A Machine Learning Method for Real-Time Analysis of Large-Scale Neural Recordings

How can we efficiently train data-constrained recurrent neural networks (dRNNs) to interpret and control large neural populations based on extensive neural recording data? This paper introduces "Convex Optimization of Recurrent Neural Networks (CORNN)", a method designed to address the inefficiencies and scalability issues of existing dRNN training algorithms. CORNN dramatically accelerates the training process—achieving speeds about 100 times faster than traditional methods—while maintaining or improving modeling accuracy. Tested on simulations involving thousands of neurons, CORNN effectively handled various computational tasks and proved robust against model mismatches and subsampling challenges. By enabling the rapid training of dRNNs with millions of parameters, CORNN represents a significant advancement in real-time neural network modeling and offers a potent tool for understanding and manipulating neural computation.

➡️ This AI Paper Introduces Relax: A Compiler Abstraction for Optimizing End-to-End Dynamic Machine Learning Workloads

How can dynamic shape computations in large language models be optimized for diverse backend environments? This paper introduces "Relax", a compiler abstraction designed to enhance the deployment of dynamic machine learning workloads. Relax implements symbolic shape annotations, enabling global tracking of dynamic shape computations throughout a program. It also features a cross-level abstraction that combines computational graphs, tensor programs, and library calls into a unified representation. This facilitates cross-level optimizations. The framework, tailored for dynamic shape models, shows competitive performance with existing hand-optimized systems on various platforms. Notably, it extends the deployment capabilities of dynamic models to mobile phones, embedded devices, and web browsers.

✅ Featured AI Tools For You

🦙 Featured AI Startups

^{*We do make a small affiliate profit when you buy this product through the click link}