Benchmark of negative transfer

Negative transfer (NT) represents that introducing source domain data/knowledge decreases the learning performance in the target domain, which is a long-standing and challenging issue in transfer learning.

This repository contains codes of NT detection experiments with over 20 representative approaches on three NT-specific datasets.

Datasets

There are three NT-specific datasets with large domain shift, i.e., on one synthetic dataset, and two real-world datasets on object recognition and emotion recognition. Their statistics are summarized below:

Running the code

The code is mainly based on Python 3.8 and PyTorch 1.8.

Code files introduction:

NT_UDA/ -- NT detection experiments on 12 unsupervised domain adaptation approaches

NT_SSDA/ -- NT detection experiments on 11 semi-supervised domain adaptation approaches

NT_Noise/ -- NT detection experiments under source label noise

NT_UDA/data_synth/ -- the synthetic dataset and code to generate it

Specifically, take NT-UDA for example, the procedure is:

Initialize the networks involved in the experiments to ensure fair comparisons by
```
python save_init_model.py
```
Run the script of a particular approach on a specific dataset
```
python demo_img_dnn.py
```

Benchmark

For unsupervised TL on DomainNet and SEED datasets:

For semi-supervised TL on DomainNet and SEED datasets:

Citation

This code is corresponding to our IEEE/CAA JAS 2023 paper below:

@article{zhang2022ntl,
  title = {A Survey on Negative Transfer},
  author = {Zhang, Wen and Deng, Lingfei and Zhang, Lei and Wu, Dongrui},
  journal = {IEEE/CAA Journal of Automatica Sinica},
  year = {2023},
  volume = {10},
  number = {2},
  pages = {305--329}	
}

Please cite our paper if you like or use our work for your research, thank you!

chamwen/NT-Benchmark

Benchmark of negative transfer

Datasets

Running the code

Benchmark

Citation