Khái niệm cốt lõi
Noise Contrastive Test-Time Training (NC-TTT) is an innovative approach that leverages noise contrastive estimation to enable unsupervised adaptation of deep learning models at test time, improving their robustness to domain shifts.
Tóm tắt
The paper presents Noise Contrastive Test-Time Training (NC-TTT), a novel unsupervised test-time training method that enhances the robustness of deep learning models to domain shifts.
Key highlights:
- NC-TTT trains a discriminator to distinguish between noisy in-distribution and out-of-distribution feature maps. This allows the model to learn a proximal representation of the source domain distribution.
- At test time, the discriminator is used to guide the adaptation of the model's encoder, moving the encoded features of target samples towards the in-distribution region.
- Experiments on various test-time adaptation benchmarks, including common corruptions (CIFAR-10/100-C) and sim-to-real domain shift (VisDA-C), demonstrate the superior performance of NC-TTT compared to recent state-of-the-art approaches.
- The authors provide a principled framework for selecting the key hyperparameters of the noise contrastive estimation, guiding the design choices.
- NC-TTT is a simple yet effective method that can be easily integrated with any CNN-based model, making it a practical and versatile solution for improving model robustness.
Thống kê
"The source domain is represented by a joint distribution P(Xs, Ys), where Xs and Ys correspond to the image and labels spaces, respectively."
"Likewise, denote as P(Xt, Yt) the target domain distribution, with Xt and Yt as the respective target images and labels."
"Following previous research, we consider the likelihood shift between source and target datasets, expressed as P(Xs|Ys) ≠ P(Xt|Yt), and assume the label space to be the same between domains (Ys = Yt)."
Trích dẫn
"NC-TTT is a simple yet effective method that can be easily integrated with any CNN-based model, making it a practical and versatile solution for improving model robustness."