A new study isolates how quantization interacts with recurrent neural network memory, introducing the concept of recurrent-state write-back to describe how a quantized state, once stored, alters all subsequent computation. Using a compact GRU model for fluorescence lifetime imaging, the researchers found that replacing continuous state propagation with deterministic 4-bit storage increased estimation errors by roughly 70x and 300x for two key parameters, caused by small updates falling below the quantization write threshold. The paper shows that techniques including error feedback, residual memory, and direction memory can recover accuracy without retraining, and confirms the same failure and recovery pattern in an independently trained LSTM.
