As the quantum counterparts to the classical artificial neural networks underlying widespread machine-learning applications, unitary-based quantum neural networks are active in various fields of quantum computation. Despite the potential, their developments have been hampered by the elevated cost of optimizations and difficulty in realizations. Here, we propose a quantum neural network in the form of fermion models whose physical properties, such as the local density of states and conditional conductance, serve as outputs, and establish an efficient optimization comparable to back-propagation. In addition to competitive accuracy on challenging classical machine-learning benchmarks, our fermion quantum neural network performs machine learning on quantum systems with high precision and without preprocessing. The quantum nature also brings various other advantages, e.g., quantum correlations entitle networks with more general and local connectivity facilitating numerical simulations and experimental realizations, as well as novel perspectives to address the vanishing gradient problem long plaguing deep networks. We also demonstrate the applications of our quantum toolbox, such as quantum-entanglement analysis, for interpretable machine learning, including training dynamics, decision logic flow, and criteria formulation.