mlsanityData to 7 Oct 2026

FORGE: Forward-Only Test-Time Adaptation for Integer-Only Vision Models on Microcontrollers

arXiv 2609.01683 ↗First posted 1 Sep 2026cs.CV cs.AR cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Vision models deployed on microcontrollers (MCUs) are quantized to integer-only arithmetic and run in inference-only runtimes that do not carry the machinery backpropagation needs: the standard tool for adapting a model to the distribution shift (sensor noise, blur, lighting) it meets in the field. Existing forward-only test-time adaptation (TTA) methods either run only on server- or edge-GPU-class models (not true microcontroller integer execution), or require the batch-normalization (BN) layers that integer deployment fuses away. We present a forward-only TTA method that operates on deployed, BN-folded, integer-only convolutional networks. The key observation is that fusing BN into the preceding convolution, a mandatory step for integer inference, destroys the statistics that normalization-based adaptation relies on. We restore adaptation by re-normalizing each folded convolution's per-channel output to its clean training statistics, using only forward-pass estimates. The method (i) recovers most of gradient-based TENT's accuracy gain (+20.9 vs. +24.9 points) and matches forward-only BN adaptation, while being the only method that runs on a folded integer-only model; (ii) needs to adapt only 3 of 21 layers (selected without seeing the test corruptions) to recover 93% of the benefit; (iii) survives single-sample streaming with a batch-size-scaled momentum; and (iv) generalizes across three datasets (up to 200 classes) and two architectures. We validate bit-exact int8 convolution execution and deploy on an ESP32-S3, where, measured with a Nordic PPK2 power profiler, the forward-only adaptation (a lightweight fp32 recalibration around the int8 convolutions) costs only 8.3 mJ (6.8% of inference energy) and 21.9 ms on the deployed SIMD-optimized model: forward-only adaptation is cheap on a real microcontroller.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 28 of its 38 references that are in the map. First posted on arXiv 1 Sep 2026.

Builds on

28 of 38 references.

References · 28

  1. 2015Deep Residual Learning for Image Recognitioncited by 51,809
  2. 2015Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shiftcited by 6,979
  3. 2018MobileNetV2: Inverted Residuals and Linear Bottleneckscited by 4,773
  4. 2019Benchmarking Neural Network Robustness to Common Corruptions and Perturbationscited by 2,857
  5. 2019Searching for MobileNetV3cited by 1,984
  6. 2017Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inferencecited by 1,548
  7. 2020Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptationcited by 844
  8. 2019Learned Step Size Quantizationcited by 604
  9. 2022Continual Test-Time Domain Adaptationcited by 520
  10. 2018Quantizing deep convolutional networks for efficient inference: A whitepapercited by 507
  11. 2022Efficient Test-Time Model Adaptation without Forgettingcited by 467
  12. 2021A White Paper on Neural Network Quantizationcited by 453
16 more
  1. 2020Improving robustness against common corruptions by covariate shift adaptationcited by 427
  2. 2023Towards Stable Test-Time Adaptation in Dynamic Wild Worldcited by 409
  3. 2021MEMO: Test Time Robustness via Adaptation and Augmentationcited by 376
  4. 2019Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Comingcited by 332
  5. 2020MCUNet: Tiny Deep Learning on IoT Devicescited by 269
  6. 2020Evaluating Prediction-Time Batch Normalization for Robustness under Covariate Shiftcited by 235
  7. 2020TensorFlow Lite Micro: Embedded Machine Learning on TinyML Systemscited by 161
  8. 2022On-Device Training Under 256KB Memorycited by 158
  9. 2022NOTE: Robust Continual Test-time Adaptation Against Temporal Correlationcited by 150
  10. 2018CMSIS-NN: Efficient Neural Network Kernels for Arm Cortex-M CPUscited by 116
  11. 2021MCUNetV2: Memory-Efficient Patch-based Inference for Tiny Deep Learningcited by 87
  12. 2024Test-Time Model Adaptation with Only Forward Passescited by 74
  13. 2025Test-Time Model Adaptation for Quantized Neural Networkscited by 9
  14. 2024On-Device Training of Fully Quantized Deep Neural Networks on Cortex-M Microcontrollerscited by 5
  15. 2025LeanTTA: A Backpropagation-Free and Stateless Approach to Quantized Test-Time Adaptation on Edge Devicescited by 1
  16. 2026Subspace Optimization for Backpropagation-Free Continual Test-Time Adaptationcited by 1

Built on by

Not cited yet.