Quantization is widely used to reduce the computational and memory cost of neural‑network inference, yet in recurrent networks the quantized state is returned at the next time step, so the rule used to store that state can alter later computations. We refer to this rule as “recurrent‑state write‑back”.
In this work we study a compact GRU encoder‑decoder designed for fluorescence‑lifetime imaging, a modality that requires estimating a short‑lived component $\tau_1$ and a long‑lived component $\tau_2$ from high‑noise time‑resolved signals.
Keeping the trained model fixed, we replace continuous state propagation with deterministic 4‑bit state storage. This change inflates the estimation error of $\tau_1$ by roughly 70× and that of $\tau_2$ by about 300×. The failure stems from many tiny updates staying below the write threshold, leaving the stored state almost unchanged while the network continues to propose changes.
Introducing error feedback, residual memory, and direction memory carries the information from these suppressed updates across time, restoring accuracy without any retraining.
Precision sweeps reveal that increasing state precision can actually degrade a fixed recurrent solution, whereas training with a matched state interface allows the network to learn compatibility.
To test generality we repeat the post‑training intervention on an independently trained LSTM. Coarse write‑back reproduces the failure, error feedback recovers accuracy, and state‑specific interventions show that the cell state is more sensitive to write‑back than the hidden state.
Overall, recurrent‑state write‑back is a key determinant of low‑precision recurrent dynamics, and the state‑storage interface should be a central design consideration for quantized recurrent inference. Review