This research brief proposes ZOA-Q, a zero-order optimization method for test-time adaptation of quantized neural networks on edge devices. Unlike existing TTA methods that assume FP32 models, ZOA-Q uses Simultaneous Perturbation Stochastic Approximation (SPSA) to estimate gradients through quantization functions, requiring only two forward passes and no backpropagation. The method adapts only batch normalization layers to maintain efficiency. The proposed 4-week sprint plan includes evaluation on ImageNet-C with expected accuracy improvements from 42.1% to 52.5% over source-only quantized models.
Key findings
Existing TTA methods fail on quantized models because discrete constraints break gradient-based optimization assumptions
ZOA-Q applies SPSA for quantization-aware gradient estimation using only two forward passes and no backpropagation
Adapting only batch normalization layers provides sufficient capacity while maintaining memory efficiency for edge deployment
Expected results show 10.4% accuracy improvement on ImageNet-C over source-only quantization with minimal latency overhead
Limitations & open questions
Proposed method lacks empirical validation as this is a research plan document
Batch normalization-only updates may provide insufficient adaptation capacity for severe domain shifts
SPSA gradient estimation may suffer from high variance with aggressive INT4 quantization