跳转至

文章背景与核心概要

模型量化是降低神经网络推理计算与内存开销的常用技术。然而在循环神经网络（RNN）中，量化后的状态会被存储并传递到后续的时间步，这意味着用于写入和存储该状态的底层机制将直接影响下游计算。本文深入探讨了循环神经网络在低精度状态存储下面临的独特挑战，并提出了“循环状态回写（recurrent-state write-back）”这一关键概念。

通过聚焦于应用于荧光寿命成像（一种定量生物成像模态）的紧凑型门控循环单元（GRU）编码器-解码器，作者剖析了低精度状态存储如何对整体性能产生严重影响。研究发现，当连续的微小更新落入写入阈值以下时，会导致存储状态几乎保持不变，而网络却在不断提出修改，从而引发严重的性能崩溃。不过，通过采用误差反馈（error feedback）、残差记忆（residual memory）和方向记忆（direction memory）等缓解策略，可以在无需重新训练模型的情况下恢复准确率。该研究表明，状态存储接口是低精度时间序列推理中不可忽视的核心设计考量。

When Quantization Breaks Memory: Recurrent-State Write-Back in Low-Precision Temporal Inference

Authors: Ismail Erbas, Xavier Intes, Vikas Pandey
Published: September 3, 2026
arXiv: 2609.04490 [cs.AI]

Authors: Ismail Erbas, Xavier Intes, Vikas Pandey
Published: September 3, 2026
arXiv: 2609.04490 [cs.AI]

Summary

Quantization is a popular technique for lowering the computational and memory overhead of neural network inference. However, in recurrent neural networks (RNNs), the quantized state is stored and passed on to the subsequent time step, meaning the mechanism used to write and store that state directly influences downstream computations.

This paper introduces the concept of recurrent-state write-back to analyze this phenomenon. Focusing on a compact Gated Recurrent Unit (GRU) encoder–decoder applied to fluorescence lifetime imaging (a quantitative biological imaging modality), the authors isolate how low-precision state storage impacts performance:

  • Severe Performance Degradation: Replacing continuous state propagation with deterministic 4-bit state storage increases estimation errors for short-lived (\(\tau_1\)) and long-lived (\(\tau_2\)) lifetime components by approximately 70× and 300×, respectively.
  • The Root Cause: Failures happen when successive tiny updates fall below the write threshold. This leaves the stored state nearly unchanged while the network continuously proposes modifications.
  • Mitigation Strategies: Techniques such as error feedback, residual memory, and direction memory successfully propagate information from suppressed updates across time steps, recovering accuracy without requiring model retraining.
  • Broader Applicability: Precision sweeps reveal that increasing state precision can occasionally degrade a fixed recurrent solution, while matched training demonstrates that models can learn compatibility with the state interface. Extending the intervention to an independently trained Long Short-Term Memory (LSTM) network shows similar vulnerabilities, with cell states demonstrating higher sensitivity than hidden states.

Ultimately, this work highlights recurrent-state write-back as a crucial factor in low-precision recurrent dynamics, proving that the state-storage interface is a primary design consideration for quantized temporal inference.

Summary

Quantization is a popular technique for lowering the computational and memory overhead of neural network inference. However, in recurrent neural networks (RNNs), the quantized state is stored and passed on to the subsequent time step, meaning the mechanism used to write and store that state directly influences downstream computations.

This paper introduces the concept of recurrent-state write-back to analyze this phenomenon. Focusing on a compact Gated Recurrent Unit (GRU) encoder–decoder applied to fluorescence lifetime imaging (a quantitative biological imaging modality), the authors isolate how low-precision state storage impacts performance:

  • Severe Performance Degradation: Replacing continuous state propagation with deterministic 4-bit state storage increases estimation errors for short-lived (\(\tau_1\)) and long-lived (\(\tau_2\)) lifetime components by approximately 70× and 300×, respectively.
  • The Root Cause: Failures happen when successive tiny updates fall below the write threshold. This leaves the stored state nearly unchanged while the network continuously proposes modifications.
  • Mitigation Strategies: Techniques such as error feedback, residual memory, and direction memory successfully propagate information from suppressed updates across time steps, recovering accuracy without requiring model retraining.
  • Broader Applicability: Precision sweeps reveal that increasing state precision can occasionally degrade a fixed recurrent solution, while matched training demonstrates that models can learn compatibility with the state interface. Extending the intervention to an independently trained Long Short-Term Memory (LSTM) network shows similar vulnerabilities, with cell states demonstrating higher sensitivity than hidden states.

Ultimately, this work highlights recurrent-state write-back as a crucial factor in low-precision recurrent dynamics, proving that the state-storage interface is a primary design consideration for quantized temporal inference.

Article Details

Article Details

Full-Text & Resources

Full-Text & Resources

License Icon Reference: license icon

License Icon Reference: license icon