VDLF-Net: Variational Feature Fusion for Adaptive and Few-Shot Visual Learning

Modeling high-dimensional visual data presents two intertwined challenges: learning feature representations that adapt to complex structures, and enabling robust inference with limited labeled examples. Although deep learning models, such as Convolutional Neural Networks, excel at representation learning and variational methods, e.g., Variational Autoencoders, offer a principled framework for uncertainty quantification, combining them for both challenges remains an active modeling problem. Driven by this limitation, this paper proposes a novel variational deep learning fusion framework. Its core innovations are two synergistic components: a feature-adaptive approximation mechanism that refines representations within a probabilistic latent space, and a variationally-regularized few-shot inference strategy that enhances generalization under data scarcity. On CIFAR-100, VDLF-Net achieves 58.34% accuracy, compared with 56.16% for ResNet-50 Enhanced. On Mini-ImageNet, its 1-shot and 5-shot accuracies are 70.72% and 86.51%, respectively, compared with 67.02% and 83.21% for GLAP. The framework couples multiscale feature fusion with variational inference within a shared training objective; the reported gains concern the evaluated benchmark settings.