Explore advanced techniques for learning with small datasets in this 25-minute conference talk from KDD 2020. Dive into practical examples, including predicting ball trajectories and global temperatures, to understand how to combine machine learning models with domain knowledge. Learn about residual modeling, regularization techniques, and methods for incorporating energy conservation principles. Discover strategies for pre-processing, post-processing, and utilizing unlabeled data to enhance model performance when working with limited training samples. Gain insights from experts Huaxiu Yao, Xiaowei Jia, Vipin Kumar, and Zhenhui Li as they demonstrate how to leverage domain-specific knowledge to improve predictions and overcome the challenges of small data scenarios.
Overview
Syllabus
Intro
Predict a ball's trajectory
Combine machine learning model and domain knowledge model
Pre-processing and post-processing
Residual modeling: Predicting the gap between
Example: predict the temperature of different places on the earth?
Example: predicting the temperature
Adding residual works well!
Predicted results with different models
Regularization: Using domain knowledge as regularization
Example: Incorporating Energy Conservation
Extension: utilizing unlabeled data
Extension: Incorporating Energy Conservation
Taught by
Association for Computing Machinery (ACM)
