Mapping from Training Data to Predictions with Datamodels - Differential Privacy in Machine Learning

Explore the intricacies of machine learning predictions and data models in this Google TechTalk presented by Andrew Ilyas. Delve into the anatomy of machine learning predictions, understanding datamodels and their role in the ML pipeline. Learn about model output functions, data generation techniques, and the construction of linear models and training sets. Discover how to construct and evaluate data models, and explore their applications in addressing model brittleness and data counterfactuals. Examine example embedding, natural similarity metrics, and spectral clustering. Gain valuable insights and takeaways from this comprehensive exploration of mapping training data to predictions using datamodels.

Syllabus

Introduction
Anatomy of a Machine Learning Prediction
Datamodels
Machine Learning Pipeline
Model Output Function
Generating Data
Linear Model
Training Set
Constructing Data Models
Evaluating Data Models
Applications
Model Brittleness
Data Counterfactuals
Example Embedding
Natural Similarity Metric
Spectral Clustering
Takeaways