LoGA: Loop-Carried Graph Attention Based Looped Vision Transformers for Fine-Grained Plant Disease Classification

Fine-grained plant disease classification requires separating classes of the same crop that often differ only in small, sparse lesion patterns. Looped Vision Transformers (LoopViTs) obtain depth by applying a small stack of weight-tied blocks repeatedly, yet each block rebuilds its attention graph from scratch at every iteration, although the same block has just computed a graph over the same tokens. We propose LoGA, a looped vision transformer with Loop-Carried Graph Attention, in which the attention graph of each block at iteration $t$ is biased by the graph the same block produced at iteration $t-1$, used as a tempered prior. In its log-prob form the update $A_t \propto \exp(QK^\top/\sqrt{d})\odot A_{t-1}^{\gamma}$ is, when the floor is inactive, equivalent to attending with a geometrically discounted sum of past query-key logits; a flash-compatible qk variant carries one step of raw logits. LoGA adds 24 scalars to a 9.91\,M-parameter model and reduces exactly to vanilla LoopViT at initialisation. We also study a parameter-free graph-stability signal $\Delta_t$ and show that, under label smoothing, predictive entropy stays near 0.671 nats, so a common entropy-exit threshold is never reached. Trained from scratch, LoGA-logprob reaches 99.78%, 94.2%, 98.6% and 99.53% top-1 accuracy on the held-out splits of PlantVillage, PlantDoc, Cassava and Plant Pathology 2020, versus 98.11%, 88.3%, 84.4% and 91.7% for vanilla LoopViT. On PlantVillage, the CLS-attention graph contracts across iterations and accuracy is maintained at twice the training depth. We regard these results as encouraging first evidence and discuss their limitations.