Speaker
Description
High-dimensional correlations present a monumental challenge in both quantum physics and deep learning, typically requiring exponential computational resources or excessively deep architectures to resolve. This talk highlights low-rank tensor structures as a unifying solution to conquer complexity across these two distinct domains. First, we pivot to highly frustrated magnetism (arXiv:2606.31021), where symmetry-optimized infinite projected entangled-pair states (iPEPS) up to D=7 are used to solve the J₁-J₂ triangular Heisenberg model. Our simulation reveals that the 120 Neel state transits to a quantum spin liquid (QSL) phase at J₂/J₁ ≈ 0.08 and the QSL precludes Z₂ gapped nature, suggesting a U(1) Dirac spin liquid behavior. Secondly, we present the Tensor-Augmented CNN (TACNN; arXiv:2604.08072), which utilizes generic tensor kernels to map data into a virtual Hilbert space, showing very competitive accuracies in comparison with other advanced CNN architecture. Together, these works demonstrate the dual power of tensors as both an elegant variational tool for quantum ground states and an efficient architectural blueprint for deep learning.