Inside TensorFlow- Building ML Infra

Explore the intricacies of building machine learning infrastructure in this 43-minute video from the Inside TensorFlow series. Join Software Engineer Mingsheng Hong as he delves into research and engineering challenges in ML infrastructure development. Discover the similarities between big data and ML infrastructure, understand the importance of investing in ML infrastructure, and examine a case study on building a new TensorFlow runtime. Learn about ML programs as computational graphs, vectorized normalization, and Eager execution. Compare ML infrastructure to SQL query processing, explore input pipelines, and understand graph processing workflows. Dive into topics such as graph rewrites, cost models, data statistics, constraint propagation, and storage/access optimizations. Gain insights into distributed and parallel execution, and recognize how ML infrastructure relates to data infrastructure with unique twists. Access additional resources through provided links and consider potential collaborations in this field.

Intro
Big data and ML infra are similar
Speaker background
Why invest in ML infra?
Case study: Building a new TF runtime
ML program as a computational graph
An example ML program
Lifetime of an ML program
Vectorized normalization
A slight digression on Eager execution
ML infra and SQL query processing
(Random) scan-based access patterns
Beyond pure dataflow
ML and DB terminology mapping
Recall graph processing workflow
Expressing input pipelines
Decoupled API and execution
Challenge: Randomized transformations
Graph rewrites
Cost model and data stats
Constraint propagation
Storage/access optimizations
Push vs pull based execution
Distributed and parallel execution
ML infra is like data infra, with new twists
Let's collaborate