Talaria - Interactively Optimizing Machine Learning Models for Efficient Inference

Explore a cutting-edge system for optimizing on-device machine learning models in this conference talk from CHI 2024. Learn about Talaria, a visualization and optimization tool designed to help practitioners create efficient ML models for personal devices. Discover how Talaria enables interactive visualization of model statistics, compilation to hardware, and simulation of optimizations to test impact on inference metrics. Gain insights from the system's two-year internal deployment, including log analysis of over 800 practitioners and 3,600 models, a usability survey assessing 20 features, and qualitative interviews with active users. Understand the challenges of balancing hardware metrics like model size, latency, and power while protecting user privacy and enabling intelligent user experiences on devices with limited resources.