Abstract
<title>Abstract</title> <p>Recent advances in Convolutional Neural Networks (CNNs) have been based on increasing depth, width, or cardinality to enhance network representational capacity. Moreover, manually designed attention modules, including Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM), exploit channel and spatial dependencies by dynamically reweighting response features. However, they suffer from additional parameters and computational overhead and rely on handcrafted importance weighting. This work proposes the SCB (Symbiotic Convolution Block), a lightweight, non-attentive module for feed-forward CNNs. SCB encourages feature diversification and recalibration by fusing shared and mutated convolutional maps across blocks, inspired by symbiotic and evolutionary principles, without explicitly learning attention maps or gating scores. SCB is a general building block that can be plugged in as the first layer of other CNN architectures with minimal overhead. It is fully end-to-end trainable on the base network. Extensive studies on CIFAR-100 with a MobileNetV1 backbone, a PlantCity dataset, and an eye disease recognition dataset, the SCB demonstrates competitive or improved classification accuracy across natural scene, plant identification, and medical image recognition tasks compared to established attention-based baselines, while being competitive in inference efficiency. These findings show implicit diversification is a potential alternative to explicit attention mechanisms in resource-constrained settings.</p>