Iterative Teacher-Aware Learning

Paper

arxiv, NeurIPS 2021, OpenReview

@misc{yuan2021iterative,
      title={Iterative Teacher-Aware Learning}, 
      author={Luyao Yuan and Dongruo Zhou and Junhong Shen and Jingdong Gao and Jeffrey L. Chen and Quanquan Gu and Ying Nian Wu and Song-Chun Zhu},
      year={2021},
      eprint={2110.00137},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Usage

To collect data and plots for a cooperative or adversarial teacher for a particular experiment, run with the command line

python3 plot_band.py -s setting_name 

where setting_name is specified by ${experiment}_${type of teacher}. Detailed settings are described in the Settings section.
And an example command to collect data and plots for Linear Classifiers on MNIST Dataset with a cooperative teacher is

python3 plot_band.py -s mnist_coop 

To use the main.py or main_irl.py script, run with the command line

python3 main.py detailed_setting_name random_seed

where detailed_setting_name is specified by ${experiment}_${type of teacher}_${mode of teacher} followed by '_' and the imitate teacher's data dimension(MNIST, Equation) or data type(CIFAR).
An example command to collect data for Linear Classifiers on MNIST Dataset with a cooperative imitate teacher with data dimension 20 and random seed 0 is

python3 main.py mnist_coop_imit_20 0

The results will be saved as pdf in the corresponding folder inside the Experiments folder.

Settings

The ${experiment} part of a setting is specified by the table

Experiment	command line
Linear Regression on Synthesized Data	regression
Linear Classifiers on Synthesized Data	class10
Linear Classifiers on MNIST Dataset	mnist
Linear Classifiers on CIFAR-10	cifar
Linear Regression for Equation Simplification	equation
Online Inverse Reinforcement Learning	irlH/irlE

The imitate teacher's data dimension(MNIST, Equation) or data type(CIFAR) part of a setting is specified by the table

Experiment	Imitate Setting 1	Imitate Setting 2
mnist	20	30
equation	40	50
cifar	9	12

The ${type of teacher} part of a setting is specified by the table

Type of Teacher	command line
cooperative	coop
adversarial	adv

The ${mode of teacher} part of a setting is specified by the table

Mode of Teacher	command line
imitate	imit

Versions

In this project, we used the following version of libraries:

Tensorflow v1.15
scikit-learn v0.22.1
numpy v1.18.2

Human Study Repo

https://github.com/yuanluya/CL_Human