Memory-Augmented GRUs for Food Loss Forecasting: A Survey of Architectures and Design Trade-offs

Abstract

Forecasting food loss is pivotal to sustainable, efficient food systems, yet it demands models that can retain long-horizon patterns, handle exogenous drivers, and remain interpretable to practitioners. This survey investigates modern time-series methods through the lens of Gated Recurrent Units (GRUs) and argues that augmenting GRUs with explicit memory is a promising, underutilized strategy for food-loss prediction. We synthesize evidence across GRU-centered variants and five families of neural memory architectures - Neural Turing Machines (NTM), Differentiable Neural Computers (DNC), Relational Memory (RelMem), Entity-centric Memory (EntMem), and Modular Memory Units (MMUs) - organizing their design choices (controllers, read/write addressing, integration points) and reported trade-offs. Drawing on comparative studies, we outline a practical MAGRU blueprint that maps memory mechanisms to common properties of food-loss series (seasonality, shocks, missingness), and distill guidance on when external memory helps, how to balance accuracy with computational cost, and how to couple memory with attention and exogenous-variable embeddings for greater interpretability. Our review identifies key gaps - limited evaluations on realistic supply-chain datasets, the absence of standard benchmarks, and little exploration of lightweight or hybrid MAGRU designs, and proposes a research agenda to close them. Overall, memory augmentation consistently strengthens long-range dependency modeling in GRU-based forecasters while introducing complexity that must be managed, positioning MAGRU as a compelling foundation for actionable, trustworthy food-loss forecasting.