Finding Generalization Measures by Contrasting Signal and Noise

Authors: Jiaye Teng, Bohang Zhang, Ruichen Li, Haowei He, Yequan Wang, Yan Tian, Yang Yuan

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, REF Complexity monotonically relates to test accuracy in real-world datasets without accessing additional validation sets, achieving 0.988 correlation on CIFAR-10 and 0.960 correlation on CIFAR-100. We further theoretically verify the utility of REF Complexity under three different cases...
Researcher Affiliation Collaboration 1Institute for Interdisciplinary Information Sciences, Tsinghua University, Beijing, China 2Peking University 3Beijing Academy of Artificial Intelligence, Beijing, China 4Shanghai Artificial Intelligence Laboratory 5Shanghai Qi Zhi Institute.
Pseudocode Yes We summarize the algorithm in Algorithm 1, which returns the REF Complexity value T β n (D, At).
Open Source Code Yes The code is available at https://github.com/962086838/REF-complexity.
Open Datasets Yes Specifically, we conduct over one hundred experiments on CIFAR-10 and CIFAR-100, and plot each regime s test accuracy and REF Complexity in Figure 1.
Dataset Splits No The paper states it achieves its results 'without accessing additional validation sets' and discusses why traditional validation sets are unnecessary. Therefore, it does not provide validation dataset splits.
Hardware Specification Yes We run all the experiments on a RTX2080Ti graphic card.
Software Dependencies No The paper mentions using SGD, but does not specify any software libraries or their version numbers (e.g., PyTorch, TensorFlow, Python versions) used in the experiments.
Experiment Setup Yes We train each model for 150 epochs. To evaluate REF Complexity correctly, each noise training process is trained five times... Hyperparameters: learning rate 0.1, 0.01, 0.001; batch size 256, 512, 1024; weight decay 1e-5, 1e-6, 1e-7; architecture Res Net20, Res Net32, Res Net44, Res Net56.