Test-Time Adaptation with Perturbation Consistency Learning (2304.12764v1)

Abstract: Currently, pre-trained LLMs (PLMs) do not cope well with the distribution shift problem, resulting in models trained on the training set failing in real test scenarios. To address this problem, the test-time adaptation (TTA) shows great potential, which updates model parameters to suit the test data at the testing time. Existing TTA methods rely on well-designed auxiliary tasks or self-training strategies based on pseudo-label. However, these methods do not achieve good trade-offs regarding performance gains and computational costs. To obtain some insights into such a dilemma, we take two representative TTA methods, i.e., Tent and OIL, for exploration and find that stable prediction is the key to achieving a good balance. Accordingly, in this paper, we propose perturbation consistency learning (PCL), a simple test-time adaptation method to promote the model to make stable predictions for samples with distribution shifts. Extensive experiments on adversarial robustness and cross-lingual transferring demonstrate that our method can achieve higher or comparable performance with less inference time over strong PLM backbones and previous state-of-the-art TTA methods.

• The paper introduces PCL as a novel test-time adaptation method that improves model robustness by disabling dropout and enforcing consistency through feature perturbations.
• The method applies consistency regularization via random perturbations and a KL-divergence objective to align outputs between original and perturbed features.
• Experimental results on diverse datasets show that PCL outperforms existing methods in accuracy and computational efficiency while maintaining stable predictions.

The paper "Test-Time Adaptation with Perturbation Consistency Learning" addresses the distribution shift problem that affects the performance of Pre-trained LLMs (PLMs) when they encounter test data distributions that differ from the training data distribution. To tackle this challenge, the authors propose a method for test-time adaptation (TTA) named Perturbation Consistency Learning (PCL).

Problem and Context:

The research highlights that existing TTA methods—like Tent and OIL—struggle to find a balance between performance improvements and computational costs. Tent uses dropout, leading to unstable predictions due to soft labels, while OIL, which employs a teacher-student paradigm, stabilizes predictions at the cost of inference speed. PCL is introduced as a solution to achieve stable predictions by turning off dropout during the test phase and encouraging consistent predictions through feature perturbations.

Methodology:

PCL enhances model robustness to distribution shifts through the following key components:

1. Turning off Dropout: By disabling dropout during test-time, the model can generate stable predictions, mitigating the risk of learning from poor-quality pseudo-labels.
2. Consistency Regularization via Perturbations: PCL introduces perturbations in the feature space to maintain robustness. This is accomplished by randomly perturbing the features and making the model’s output prediction consistent for both original and perturbed features. This perturbation is achieved using both dropout and Gaussian noise.
3. KL-Divergence Objective: The learning objective is defined using Kullback-Leibler divergence to ensure that the predictions of perturbed features are similar to those of the original features.

Experimental Evaluation:

The paper presents extensive experiments on multiple datasets, including NoiseQA (with synthetic and natural noise), XQuAD, MLQA for question answering tasks, and RockNER for named entity recognition tasks. The proposed PCL method is benchmarked against strong baseline methods such as direct fine-tuning, xTune, Tent, EATA, and OIL.

Key results indicate that:

• PCL outperforms Tent and EATA in both effectiveness and computational efficiency, offering higher accuracy with improved inference times.
• PCL achieves performance similar to OIL but with significantly reduced computational overhead.
• In scenarios involving robustness tuning (e.g., with xTune), PCL further enhances performance.

Findings and Contributions:

Overall, the research provides several significant insights and contributions:

• It identifies the limitations of typical fully test-time adaptation methods in NLP, particularly due to unstable outputs caused by dropout.
• PCL method is proposed as an effective alternative, balancing model robustness against distribution shifts with computational efficiency.
• The experimental analysis demonstrates that PCL achieves state-of-the-art results across various datasets and settings, including adversarial attacks and cross-lingual transfer tasks.

The paper contributes to advancing the reliability and robustness of NLP models by enabling adaptive test-time learning with minimal computational burden, addressing a critical challenge in deploying PLMs in real-world, dynamic environments.