Explore data efficiency through transfer learning in this 50-minute conference talk by Eddie Du from Open Data Science. Learn how to apply cutting-edge academic research in transfer learning to real-world business problems, including the cold-start issue. Discover a hybrid instance-based transfer learning approach that outperforms baselines and uses probabilistic weighting to fuse information from source to target domains. Examine a framework for building differentially private aggregation approaches to transfer knowledge from existing models to new companies with limited data. Understand how these methods can increase customer trust and advance revenue. Delve into topics such as data aggregation, binary classification while preserving privacy, differential privacy in machine learning, and the effects of epsilon on performance.
Data Efficiency Through Transfer Learning - Eddie Du
Overview
Syllabus
Intro
About Georgian Partners
Proposed Solution: Data Aggregation with Transfer Learning
Instance-based Transfer Learning
Mathematical Justification
Bluecore's Challenge: Binary classification while preserving privacy
Proposed Solution: Aggregate at the Model level
What does it mean to preserve privacy?
First attempt at preserving privacy
Revisiting the Compensation Example
Quantifying Privacy with Differential Privacy
Differential Privacy in Machine Learning
Using Differential Privacy in Practice
Recall the Proposed Solution for Bluecore
Differentially Private Logistic Regression
Effect of Epsilon on Performance
Summary
Taught by
Open Data Science
