Dive into an in-depth exploration of attention mechanisms with linear biases in this 39-minute video. Uncover the innovative Attention with Linear Biases (ALiBi) method, which introduces a novel position representation technique in transformer models. Learn how ALiBi penalizes query-key attention scores proportionally to their distance without adding explicit positional embeddings, enabling efficient extrapolation to longer sequence lengths beyond those seen during training. Explore the research paper "Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation" by Ofir Press, Noah A. Smith, and Mike Lewis. Gain insights into the latest AI research and industry trends through recommended resources such as The Deep Dive newsletter and Unify's blog. Connect with the Unify community through their website, GitHub, Discord, and Twitter to further engage with AI deployment stack discussions and developments.
Overview
Syllabus
Attention with Linear Biases Explained
Taught by
Unify