MEMO: Test Time Robustness via Adaptation and Augmentation

By Marvin Zhang, Sergey Levine, Chelsea Finn. Neural Information Processing Systems, 2021

543 citations recorded by the source corpus.

While deep neural networks can attain good accuracy on in-distribution test points, many applications require robustness even in the face of unexpected perturbations in the input, changes in the domain, or other sources of distribution shift. We study the problem of test time robustification, i.e., using the test input to improve model robustness. Recent prior works have proposed methods for test time adaptation, however, they each introduce additional assumptions, such as access to multiple test points, that prevent widespread adoption. In this work, we aim to study and devise methods that make no assumptions about the model training process and are broadly applicable at test time. We propose a simple approach that can be used in any test setting where the model is probabilistic and adaptable: when presented with a test example, perform different data augmentations on the data point, and then adapt (all of) the model parameters by minimizing the entropy of the model's average, or marginal, output distribution across the augmentations. Intuitively, this objective encourages the model to make the same prediction across different augmentations, thus enforcing the invariances encoded in these augmentations, while also maintaining confidence in its predictions. In our experiments, we evaluate two baseline ResNet models, two robust ResNet-50 models, and a robust vision transformer model, and we demonstrate that this approach achieves accuracy gains of 1-8\% over standard model evaluation and also generally outperforms prior augmentation and adaptation strategies. For the setting in which only one test point is available, we achieve state-of-the-art results on the ImageNet-C, ImageNet-R, and, among ResNet-50 models, ImageNet-A distribution shift benchmarks.

Subfields: Civil and Structural Engineering, Control and Systems Engineering, Hardware and Architecture

View Paper: https://www.semanticscholar.org/paper/6e2a5c59f8119374bc68e3dab075d660700b8155

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Civil and Structural Engineering

In Control and Systems Engineering

In Hardware and Architecture

Citations in the corpus

Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Models (2022, methodology): “Different from the previous single-point method [49], we refine the entropy minimization by proposing confidence selection , which helps filter out noisy augmentations that may lead to misleading predictions.”
Domain Generalization: A Survey (2021, background): “On the other hand, TTT differs from DA in that only a single [161] ormini-batch [162] test data is used formodel tuning, which is often done in an online manner [163] and of course, without human supervision.”
Continual Test-Time Domain Adaptation (2022, background): “A number of works have also shown that data augmentation [52] during training [14, 16, 32, 69] and during testing [1, 40, 73] can improve model robustness and generalizability.”
Towards Stable Test-Time Adaptation in Dynamic Wild World (2023, methodology): “In TTA, prior methods often conduct adaptation on pre-trained models with batch normalization (BN) layers (Ioffe & Szegedy, 2015), and most of them are built upon BN statistics adaptation (Schneider et al., 2020; Nado et al., 2020; Khurana et al., 2021; Wang et al., 2021; Niu et al., 2022a; Hu et al”
Efficient Test-Time Model Adaptation without Forgetting (2022, methodology): “6) 10,000 × 32 0 • MEMO (Zhang et al., 2021b) 19.6 ( − 2 .”
A Comprehensive Survey on Test-Time Adaptation Under Distribution Shifts (2023, unknown): “These methods are generally divided into two main categories: (1) training with auxiliary tasks [8], [365], [390], which introduces an additional selfsupervised learning task in the primary task during both training and test phases, and (2) fine-tuning with unsupervised objectives [14], [391], [392]”