Understanding the Generalization Benefit of Model Invariance from a Data Perspective

Authors: Sicheng Zhu, Bang An, Furong Huang

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments on multiple datasets, we evaluate sample covering numbers for some commonly used transformations and show that the smaller sample covering number for a set of transformations (e.g., the 3D-view transformation) indicates a smaller gap between the test and training error for invariant models, which verifies our propositions.
Researcher Affiliation Academia Department of Computer Science University of Maryland, College Park {sczhu, bangan, furongh}@umd.edu
Pseudocode Yes In experiments, we use modified k-medoids [35] clustering method to find the approximation of N(ϵ, S, ρG) (see Algorithm 1). ... Algorithm 1 Sample Covering Number Estimation
Open Source Code Yes Code is available at https://github.com/bangann/understanding-invariance.
Open Datasets Yes We report experimental results on CIFAR-10 [29] and Shape Net [10] in this section
Dataset Splits No The paper describes sampling training images and reports results for different sample sizes per class, but it does not specify a distinct validation set or the exact percentages/counts for train/validation/test splits.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory, or cloud instances) used to conduct the experiments.
Software Dependencies No The paper mentions using ResNet18 and methods like data augmentation and KL divergence regularization but does not list specific software libraries (e.g., PyTorch, TensorFlow) with their version numbers.
Experiment Setup Yes We use Res Net18 [25] on both datasets... A simple method to learn invariant models is to do data augmentation. The augmented loss function is Laug(x) = L(f(g(x)))... We use this method on CIFAR-10 and Shape Net... We further enforce the invariance using the invariance regularization loss similar to [48, 51]: L = Lcls(f(x)) + λKL(f(x), f(g(x))). ... Table 3 shows specific λ values.