Frequency Composition for Compressed and Domain-Adaptive Neural Networks

2025-05-27

AI Summary

Neural networks for edge devices must jointly address model compression and domain adaptation, yet these objectives have traditionally been treated in isolation. This paper proposes CoDA, the first framework to unify quantization-aware training (QAT) and test-time adaptation (TTA). CoDA leverages frequency-domain decomposition: low-frequency components underpin robust model compression, while high-frequency components encode domain-specific signals. During training, QAT is performed exclusively on the low-frequency subspace; at inference, TTA operates source-free—adapting dynamically via high-frequency–driven batch normalization without requiring source-domain data. Evaluated on CIFAR-10-C and ImageNet-C, CoDA achieves absolute accuracy improvements of +7.96% and +5.37% over full-precision TTA baselines, respectively, while enabling substantial model compression. This synergy between frequency-aware quantization and test-time adaptation establishes a new paradigm for efficient, robust deployment of neural networks on resource-constrained edge devices.

Technology Category

• Machine Learning: Learning on the Edge & Model Compression
• Computer Vision: Adversarial Attacks & Robustness
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Search and Retrieval-Augmented AI: Vertical and domain-specific search

Abstract

Modern on-device neural network applications must operate under resource constraints while adapting to unpredictable domain shifts. However, this combined challenge-model compression and domain adaptation-remains largely unaddressed, as prior work has tackled each issue in isolation: compressed networks prioritize efficiency within a fixed domain, whereas large, capable models focus on handling domain shifts. In this work, we propose CoDA, a frequency composition-based framework that unifies compression and domain adaptation. During training, CoDA employs quantization-aware training (QAT) with low-frequency components, enabling a compressed model to selectively learn robust, generalizable features. At test time, it refines the compact model in a source-free manner (i.e., test-time adaptation, TTA), leveraging the full-frequency information from incoming data to adapt to target domains while treating high-frequency components as domain-specific cues. LFC are aligned with the trained distribution, while HFC unique to the target distribution are solely utilized for batch normalization. CoDA can be integrated synergistically into existing QAT and TTA methods. CoDA is evaluated on widely used domain-shift benchmarks, including CIFAR10-C and ImageNet-C, across various model architectures. With significant compression, it achieves accuracy improvements of 7.96%p on CIFAR10-C and 5.37%p on ImageNet-C over the full-precision TTA baseline.

Problem

• Unifying model compression and domain adaptation challenges
• Enabling compressed models to learn generalizable features
• Adapting to target domains using frequency components

Innovation

• Frequency composition unifies compression and adaptation
• Quantization-aware training with low-frequency components
• Source-free test-time adaptation using full-frequency data