The Parrot: A Tutorial on Retrieval Assisted Generation

Retrieval Augmented Generation (RAG) is an architecture that integrates a generative text model with external sources not present during training time, providing an effective, nonparametric alternative to model fine tuning. This tutorial will provide an overview of all the components in RAG and how to wire them together. Additionally, we will describe some useful extensions to RAG and recent advances, such as re-ranking with cross encoders and hypothetical document embeddings.

An introduction of self-supervised learning in Image, Video, and Audio

Discover the game-changing world of self-supervised learning! This beginner-friendly talk introduces you to the exciting realm of self-supervised learning techniques for images, videos, and audio. Instead of relying on tons of manual labels, self-supervised learning leverages the structure information within data to generate useful representations. We'll delve into the basics, methods, and uses of self-supervised learning in these domains. Get ready to see how this approach supercharges deep neural network training, making it efficient and scalable. Join us to explore the world of self-supervised learning and how it's shaping the future of tackling intricate tasks in image, video, and audio.

LLMs and the Groq Language Processing Unit (LPU™)

Groq's newly announced Language Processor Unit, the Groq LPU, has demonstrated that it can run 70-billion-parameter enterprise-scale language models at a record speed of more than 230 tokens per second. Groq chips (with Dataflow inside) are optimized for the sequential nature of natural language and other sequential data like DNA, music and code. Being so specific in the design of the LPU leads to much better performance on language tasks than, for example, GPUs that are optimized for parallel graphics processing.

On the hardware front, Groq utilizes SRAM instead of a GPUs HBM, Dataflow instead of multi-core parallelism, and a kernel-less compiler from Python, instead of hand-tunes kernels. This has allowed Groq to be first at reaching over 230 Tokens / second for large language model inference at model sizes where GPUs are struggling, such as 70B or more parameters, or ultra low latency.

My favorite part is that the Groq architecture is deterministic. A Groq program runs in exactly the same number of clock cycles every time, and the results are bit identical, every time, as opposed to dynamic GPUs which have a natural variability due to dynamic scheduling of the cores and their large dynamic memories.

Experimentation for Multi-Modal AI: Insights from DataRobot

The landscape of multi-modal AI offers valuable opportunities to extract insights & signals from diverse sources like text, images, and geospatial data. This can be a particularly appealing approach for asset valuations, for example with real estate. However, addressing the complexity of these challenges requires streamlined approaches to experimentation and deployment in order to efficiently evaluate them. This session aims to equip attendees with practical insights and tools to navigate this complexity. The DataRobot presenter, well-versed in multi-modal AI and data science, will share their insights and experiences. They will demonstrate the exploration of various algorithms, preprocessing steps, and problem framings. From uncovering feature interactions to optimizing model performance, DataRobot will showcase these in an accessible manner & the team will highlight learnings from their extensive work with various financial institutions across the world.