Deep Active Learning

Python implementations of the following active learning algorithms:

• Random Sampling
• Least Confidence [1]
• Margin Sampling [1]
• Entropy Sampling [1]
• Uncertainty Sampling with Dropout Estimation [2]
• Bayesian Active Learning Disagreement [2]
• K-Means Sampling [3]
• K-Centers Greedy [3]
• Core-Set [3]
• Adversarial - Basic Iterative Method
• Adversarial - DeepFool [4]

Prerequisites

• numpy 1.14.3
• scipy 1.1.0
• pytorch 0.4.0
• torchvision 0.2.1
• scikit-learn 0.19.1
• ipdb 0.11

Usage

$ python run.py

Reference

[1] A New Active Labeling Method for Deep Learning, IJCNN, 2014

[2] Deep Bayesian Active Learning with Image Data, ICML, 2017

[3] Active Learning for Convolutional Neural Networks: A Core-Set Approach, ICLR, 2018

[4] Adversarial Active Learning for Deep Networks: a Margin Based Approach, arXiv, 2018