Deep Active Learning: Enhancing Data Efficiency in Machine Learning - Lecture

Explore deep learning's data efficiency challenges and potential solutions in this comprehensive lecture by Carnegie Mellon University's Zachary Lipton. Delve into innovative approaches for enhancing labor efficiency in human-interactive systems, including dialogue policy learning, deep active learning for NLP, and strategies for handling noisy and limited labeled data. Examine the concept of active learning with partial feedback and discover a novel method for reducing NLP models' reliance on spurious data associations. Gain insights into the intersection of machine learning, social impact, and applications in clinical medicine and natural language processing from this expert in the field.

Syllabus

Introduction
About the Lab
Credit
Deep Learning
How hungry are these systems
More bang for the data
Label shift assumptions
Debt augmentation
Noise invariant representations
Transfer learning
Active Learning Approach
Denovo Active Learning
Active Learning Example
Active Learning Questions
Traditional Acquisition Functions
Dropout Regularization
Weight Uncertainty
Objective
Context
Thompson Sampling
Uncertainty Estimates
Data Hungry Tasks
Retraining
Problems
Active Learning with Partial Feedback
Expected Information Gain
Different Steps
Crowdsourcing
Labeling
The Worker