Uncertainty is an omnipresent aspect of real-world data, necessitating innovative approaches for representation learning. In this paper, we introduce an avant-garde model, namely, the neutrosophic-based transformer network (NTN), which leverages the fusion of neutrosophic logic and transformer architecture to address the multifaceted challenge of modeling and managing uncertainties. The design of NTN includes three primary building blocks, namely, neutrosophic encoding, multipath network, and fusion and decision modules. The neutrosophic encoding module applies a convolving window to map image data into the neutrosophic domain, with three subsets, namely, truth, indeterminacy, and falsehood, which model the inherent uncertainties in pixel attributes. Then, three network paths are built with a transformer encoder to extract rich and adaptive representations from the generated neutrosophic data, making up the multipath network module. Finally, the fusion and decision modules combine diverse knowledge from the paths, enabling comprehensive representation learning. Extensive experimentations on Fashion-MNIST and CIFAR-10 dataset validate the effectiveness and efficiency of NTN, outperforming the cutting-edge vision models under different levels of uncertainties. The NTN paves the way for a new era of representation learning, where uncertainties are harnessed as a valuable resource instead of an impediment, promising broad applicability in real-world settings, in which data are intrinsically uncertain