Efficient Active Learning for Gaussian Process Classification by Error Reduction

This is the source code for the paper Efficient Active Learning for Gaussian Process Classification by Error Reduction published in Neurips2021.

As we speeded up the integral calculation with Gaussian qudrature, now the running speed of NR-SMOCU-SGD and NR-(S)MOCU-RO is even faster than the results shown in the paper.

Running

The code includes two directories corresponding to two scenarios of active learning (AL): query synthesis scenario (with continous search space) and pool-based scenario (with discrete search space). In each directory, just run LocalRunner.py to compare the performance of different active learning algorithms.

For pool-based AL problem, the recommended algorithms are NR-MOCU-RO and NR-SMOCU-RO, while for query synthesis AL problem, the recommended algorithm is NR-SMOCU-SGD.

For any questions and issues, please contact [email protected]