We explore FERN (Feature-Entangled Recurrent Network), a neural architecture that supplementsconventional convolutional feature extraction with relational operations intended to capture both functionaldependencies and co-activation patterns between feature channels. The core FERN block combines twocomplementary signals computed from shared projections: (1) a bilinear path that computes implicitpairwise channel products, which may enable polynomial approximation of functional relationships throughiteration, and (2) a channel attention path that captures global co-activation patterns across spatial positions.These blocks are applied recurrently within each resolution stage, producing computational depth throughweight-shared iteration rather than sequential layer stacking. Under a matched training setup on CIFAR-10,FERN reaches 91.90% test accuracy (mean over 3 seeds), compared to ResNet-110 at 90.97% in the samesetting (+0.93%), while using 12.4% fewer parameters (1.52M vs 1.73M). We report ablations that suggestboth relational components contribute to the result and that recurrence offers incremental improvement.These findings are preliminary: the comparison is on a single dataset under a single optimizerconfiguration, and broader validation across datasets, training recipes, and scales remains future work.
Mohammad Hassan (Mon,) studied this question.