Neural Complexity Measures

The official repository for the paper Neural Complexity Measures (NeurIPS 2020) by Yoonho Lee et al.

This work proposes a meta-learning framework for predicting generalization.

Neural Complexity (NC) is a neural network which predicts the generalization gap of other networks.

1D Regression Experiments

The code inside 1d_regression/ is orgnized as follows.

• run.py : Main entry point. Implements MemoryBank and NC's specific training loop. Run with python run.py --OPTIONS
• model/ : Contains definition of the NC network, along with parallelized task learners.
• data/ : Sinewave data generator

Citation

If you find this useful in your research, please consider citing our paper:

@misc{lee2020neural,
    title={Neural Complexity Measures},
    author={Yoonho Lee and Juho Lee and Sung Ju Hwang and Eunho Yang and Seungjin Choi},
    year={2020},
    journal={arXiv preprint arXiv:2008.02953},
}