SKILL 102 Lifelong Learning Dataset

1Thomas Lord Department of Computer Science, University of Southern California, 2Neuroscience Graduate Program, University of Southern California 3Intel Labs 4Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences 5Dornsife Department of Psychology, University of Southern California *Equal contribution as second author

102 Dataset contains 102 datasets across Fine-grained task such as Flower-102, medical image tasks, and others. It is used for Lifelong learning.

Download

            $ wget http://ilab.usc.edu/andy/skill-dataset/separate-v/skill-dataset.z01
            $ wget http://ilab.usc.edu/andy/skill-dataset/separate-v/skill-dataset.z02
            $ wget http://ilab.usc.edu/andy/skill-dataset/separate-v/skill-dataset.zip
            $ unzip skill-dataset.zip

Compared with other Lifelong learning benchmark

Main Figure

(a) SKILL-102 dataset visualization. Task difficulty (y-axis) was estimated as the error rate of a ResNet-18 trained from scratch on each task for a fixed number of epochs. Circle size reflects dataset size (number of images). (b) Comparison with other benchmark datasets including Visual Domain Decathlon Cifar-100, F-CelebA, Fine-grained 6 tasks. c) Qualitative visualization of other datasets, using the same legend and format as in a).

Results with Continuous Learning methods

Main Figure

Result of average absolute accuracy on all tasks learned so far, as a function of the number of tasks learned. The sharp decrease in early tasks carries no special meaning except for the fact that tasks 4,8,10 are significantly harder than the other tasks in the 0-10 range, given the particular numbering of tasks in SKILL-102.

Main Figure

Accuracy on task 1 (learning to classify 102 types of flowers) as a function of the number of tasks learned.

BibTeX

If you are interesting using this dataset, please consider to cite

@article{ge2023lightweight,
      title={Lightweight Learner for Shared Knowledge Lifelong Learning},
      author={Ge, Yunhao and Li, Yuecheng and Wu, Di and Xu, Ao and Jones, Adam M and Rios, Amanda Sofie and Fostiropoulos, Iordanis and Wen, Shixian and Huang, Po-Hsuan and Murdock, Zachary William and others},
      journal={arXiv preprint arXiv:2305.15591},
      year={2023}
    }
    
}

We also extend to 107 for more details on how to create:link to here