A Hybrid Neuro-Fuzzy Approach for Heterogeneous Patch Encoding in ViTs Using Contrastive Embeddings & Deep Knowledge Dispersion

Vision Transformers (ViT) are commonly utilized in image recognition and related applications. It delivers impressive results when it is pre-trained using massive volumes of data then employed mid-sized or small-scale evaluations such as ImageNet CIFAR-100. Basically, converts images into patches, the patch encoding used to produce latent embeddings (linear projection positional embedding). In this work, module modified heterogeneous embedding by new types weighted encoding. A traditional transformer uses two including linear embedding. The proposed model replaces with combination embedding, three additional called Spatial Gated, Fourier Token Mixing Multi-layer perceptron Mixture Secondly, a Divergent Knowledge Dispersion (DKD) mechanism propagate previous information far network. ensures knowledge be multi headed attention for efficient Four benchmark datasets (MNIST, Fashion-MNIST, CIFAR-10 CIFAR-100) comparative performance evaluation. named SWEKP-based ViT, where term SWEKP stands Stochastic Weighted Composition Contrastive Embeddings & Heterogeneous Patch Encoding. experimental show that adding extra integrating DKD increases datasets. ViT has been trained separately these Conclusively, spatial gated default outperforms MLP-Mixture embeddings.