Current Should Not Sneak: Constrained Codes for Reliable Memristor Crossbar Arrays

What the math gives to ML

The paper develops non-binary constrained codes that eliminate or strongly suppress short rectangular sneak paths in memristor arrays, using GF(4) codes for two simultaneously read rows and GF(8) codes for three. The transferable asset is a constructive, rate-aware representation that enforces local combinatorial constraints on stored symbols while retaining a measurable information rate. A direct ML application is reliable storage of quantized neural-network weights or embeddings in ReRAM: encode blocks before writing, decode after reading, and test whether the reduction in read errors improves model accuracy enough to justify coding overhead. The central engineering trade-off is explicit: inserting a zero row can force S_4=0 but changes the rate to R_new=2R/3.