level: research
tensor data appears in neuroimaging, genomics, and climate science, where multiway relationships are lost if you flatten the data into a vector. current methods either force a single low-rank structure or ignore the tensor shape entirely. the dual-channel tensor neural network (dc-tnn) handles this by decomposing each input into a low-rank core and a sparse refinement, then processing both through separate neural channels. the framework works with cp, tucker, or tensor-train cores without needing to pre-specify the format.
the paper provides finite-sample risk bounds for the dc-tnn estimator. the total error breaks down into network approximation error, core estimation error, and refinement selection error. the effective model complexity depends on both the core and refinement channels, not just the raw input size. this theoretical backing shows the estimator converges at a rate that adapts to the underlying tensor structure.
a key practical feature is conformal structure selection, which automatically picks the best tensor decomposition and refinement threshold using a holdout set. this avoids manual tuning and gives valid prediction intervals. experiments on synthetic and real datasets show the method outperforms single-channel tensor networks and vectorized deep learning baselines, especially when the signal is a mix of smooth and sparse components.
why it matters: it gives reliable error guarantees and automatic model selection for tensor data, making deep learning safer for high-stakes fields like medical imaging.