Recurrent-State Write-Back Can Break Low-Precision Inference

Summary

The study examines how storing recurrent neural-network states in low precision changes later computations. It introduces the term recurrent-state write-back and isolates the effect in a compact GRU encoder-decoder for fluorescence lifetime imaging, where the model estimates short- and long-lived components from noisy time-resolved signals. With the trained model held fixed, deterministic 4-bit state storage increases estimation errors for the two lifetime parameters by approximately 70 times and 300 times. The failure occurs when repeated small state updates fall below the write threshold, so the stored state remains almost unchanged even as the network continues to propose updates. Error feedback, residual memory, and direction memory preserve information from these suppressed updates and restore accuracy without retraining. Precision sweeps show that simply increasing state precision can make a fixed recurrent solution worse, while matched training can learn compatibility with the chosen state interface. A post-training intervention in an independently trained LSTM reproduces the coarse-write-back failure, and error feedback again restores accuracy. Interventions applied separately to the LSTM states indicate that the cell state is more sensitive than the hidden state. The results identify the state-storage interface, rather than nominal arithmetic precision alone, as a key design consideration for quantized recurrent inference.