TY  - JOUR
AU  - Divakar, Subham
AU  - Priyadarshini, Rojalina
PY  - 2026
DA  - 2026/10/29
TI  - LoGA: Loop-Carried Graph Attention Based Looped Vision Transformers for Fine-Grained Plant Disease Classification
JO  - ICCK Transactions on Emerging Topics in Artificial Intelligence
T2  - ICCK Transactions on Emerging Topics in Artificial Intelligence
JF  - ICCK Transactions on Emerging Topics in Artificial Intelligence
VL  - 3
IS  - 4
SP  - 218
EP  - 238
DO  - 10.62762/TETAI.2026.672726
UR  - https://www.icck.org/article/abs/TETAI.2026.672726
KW  - Fine-grained plant disease classification
KW  - looped transformers
KW  - vision transformer
KW  - attention graph
KW  - weight tying
KW  - recurrent depth
KW  - dynamic exit
AB  - 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.
SN  - 3068-6652
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Divakar2026LoGA,
  author = {Subham Divakar and Rojalina Priyadarshini},
  title = {LoGA: Loop-Carried Graph Attention Based Looped Vision Transformers for Fine-Grained Plant Disease Classification},
  journal = {ICCK Transactions on Emerging Topics in Artificial Intelligence},
  year = {2026},
  volume = {3},
  number = {4},
  pages = {218-238},
  doi = {10.62762/TETAI.2026.672726},
  url = {https://www.icck.org/article/abs/TETAI.2026.672726},
  abstract = {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.},
  keywords = {Fine-grained plant disease classification, looped transformers, vision transformer, attention graph, weight tying, recurrent depth, dynamic exit},
  issn = {3068-6652},
  publisher = {Institute of Central Computation and Knowledge}
}