dianas1108 / imt

A Unified Implementation of IMT

As part of my master's thesis, this repository contains the implementation of a series of iterative machine teaching (IMT) methods proposed in Liu's line of work. We thus also provide a fair comparison of these different methods.

Specifically, the ipynb files contained in the src folder respectively implement:

• IMT: the baseline IMT method by sample selection
• LAST: label synthesis teaching
• DHT: data hallucination teaching
• Joint: a natural extension of the above methods that jointly synthesise label and input

For each of LAST, DHT and Joint, the method can be modified to a constrained version by commenting out/uncommenting few lines of code.

Besides these greedy methods, Parametrised implements the more challenging parametrised DHT trained using unrolling.

After setting up the environment with the light dependencies specified in requirements.txt, the scripts can be easily run and will store their results in the results folder. The performance of difference methods can then be compared in plot.ipynb.

The experiments here are based on the half-moon dataset. We have generated and provided this simple demonstrative dataset in data for easy usage. The codes can be adapted to use for other datasets with modifications.

Acknowledgement

During the development of this project, we have referred to

• Data Hallucination Teaching
• Iterative Machine Teaching