Summary of Test-time Model Adaptation with Only Forward Passes, by Shuaicheng Niu et al.
Test-Time Model Adaptation with Only Forward Passes
by Shuaicheng Niu, Chunyan Miao, Guohao Chen, Pengcheng Wu, Peilin Zhao
First submitted to arxiv on: 2 Apr 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Test-time adaptation is crucial for adapting trained models to unseen test samples with potential distribution shifts. However, in real-world scenarios, models are often deployed on resource-limited devices like FPGAs and are quantized and hard-coded with non-modifiable parameters for acceleration. To address this, we propose a test-time Forward-Optimization Adaptation (FOA) method that solely learns a newly added prompt via a derivative-free covariance matrix adaptation evolution strategy. FOA utilizes a novel fitness function measuring test-training statistic discrepancy and model prediction entropy to work stably under online unsupervised settings. Moreover, an activation shifting scheme is designed to tune model activations for shifted test samples, aligning them with the source training domain. FOA outperforms gradient-based TENT on full-precision 32-bit ViT while achieving a 24-fold memory reduction on ImageNet-C. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine having a super smart AI that can adapt to new situations it’s never seen before. This is called test-time adaptation, and it’s really important for real-world applications like image recognition. The problem is that these models usually need powerful computers to work, but in reality, they’re often run on tiny devices with limited resources. Our solution is a new method called Forward-Optimization Adaptation (FOA) that helps the AI learn from new situations without needing powerful computers. FOA uses a clever combination of mathematical tricks and image processing techniques to adapt the model to new situations. This means we can use these models on devices with limited resources, like smartphones or tablets. |
Keywords
» Artificial intelligence » Optimization » Precision » Prompt » Unsupervised » Vit
