Quantization reduces the computational and memory demands of neural-network inference but can affect recurrent networks due to state storage rules. The introduced recurrent-state write-back isolates this effect in a GRU encoder--decoder used for fluorescence lifetime imaging.
Replacing continuous state propagation with deterministic 4-bit state storage increased estimation errors for key parameters by approximately 70x and 300x. Failure occurs when small updates remain below the write threshold, leaving the stored state nearly fixed while the network continues to propose changes.
Error feedback, residual memory, and direction memory can recover accuracy without retraining. Increasing state precision can worsen fixed solutions, but matched training can learn compatibility with the state interface. Similar effects are observed in an LSTM, where coarse write-back reproduces failure, and error feedback restores accuracy.
These findings establish recurrent-state write-back as a key factor in low-precision recurrent dynamics and emphasize the importance of the state-storage interface in quantized recurrent inference.
Source: https://arxiv.org/abs/2609.04490