EVENT DOSSIER
A cycle state write-back mechanism was proposed to address the memory failure issue caused by low-precision quantization. Intervention measures such as error feedback were required, and the model needed to be retrained.
Researchers found that using coarse-grained 4-bit quantization to store internal states in recurrent neural networks (GRU, LSTM) leads to the “recurrent state write-back problem,” causing the network to recalculate based on outdated memories, severely compromising the accuracy of short-term and long-term parameter estimation. In time-resolution fluorescence lifetime imaging tasks, this problem resulted in approximately 70 times and 300 times higher errors in key parameters, respectively. Experiments showed that simply increasing the bit width or changing storage rules without retraining did not solve the problem fundamentally; by introducing lightweight interventions such as error feedback, residual memory, and direction memory, most of the accuracy could be restored without retraining. Additionally, independent verifications revealed that unit states are more sensitive than hidden states, establishing the recurrent state write-back mechanism as a key determinant of low-precision recurrent dynamics. The state storage interface was identified as a core design consideration for quantized recurrent inference.
政府 · 科研机构across 1 days
A cycle state write-back mechanism was proposed to address the memory failure issue caused by low-precision quantization. Intervention measures such as error feedback were required, and the model needed to be retrained.
Researchers found that using 4-bit quantization to store internal states in GRU/LSTM leads to cycle state write-back problems, causing the error in key parameters in fluorescence lifetime imaging tasks to increase by approximately 70 times and 300 times, respectively.
2 reports研究人员发现，在循环神经网络（GRU 和 LSTM）中，使用粗粒度 4 位量化存储内部状态会导致“循环状态写回问题”，使网络在数十步内基于陈旧记忆重复计算。在时间分辨荧光寿命成像任务中，仅改变存储规则而未重新训练模型，导致两个关键参数的估计误差分别增加约 70 倍和 300 倍。研究证实，修复的关键在于学习动态与存储规则的兼容性而非单纯增加位宽；携带舍入误差、保留残差或追踪重复次阈值更新方向等轻量级干预措施均能恢复大部分精度且无需重新训练。此外，独立训练的 LSTM 验证了单元状态对此问题比隐藏状态更敏感。
The researchers proposed a “circular state write-back” mechanism to address the issue of memory impairment caused by low-precision quantization in temporal reasoning. In the GRU encoder-decoder for fluorescence lifetime imaging, replacing continuous state propagation with deterministic 4-bit state storage resulted in approximately 70 times and 300 times higher estimation errors for short-lived and long-lived parameters, respectively. When repetitive small updates fall below the write threshold, the stored states become nearly fixed while the network continues to propose changes, leading to failure; precision can be restored without re-training by using error feedback, residual memory, and direction memory to carry suppressed update information. Experiments showed that improving state precision may worsen fixed-cycle solutions, while matching training demonstrated compatibility with state interfaces; repeating this intervention in independently trained LSTMs also reproduced both failure and recovery phenomena, revealing that cell states are more sensitive than hidden states. The study established circular state write-back as a key determinant of low-precision circular dynamics and identified state storage interfaces as a core design consideration for quantized circular reasoning.