Dual-Channel Tensor Neural Networks: Finite-Sample Theory and Conformal Structure Selection
Researchers have introduced a Dual-Channel Tensor Neural Network (DC-TNN) designed to handle tensor-valued data, which is common in fields like neuroimaging and genomics. This new network decomposes tensor inputs into a low-rank core and a sparse refinement, processing them through coupled neural channels. The framework establishes non-asymptotic risk bounds for estimation and offers a structure-aware conformal procedure for inference and structure selection, demonstrating competitive accuracy and reliable uncertainty quantification on simulated and real-world datasets. AI
IMPACT Introduces a novel neural network architecture for processing complex tensor-valued data, potentially improving analysis in fields like neuroimaging and genomics.