Abstract
We propose a natural quantization of a standard neural network, where theneurons correspond to qubits and the activation functions are implemented viaquantum gates and measurements. The simplest quantized neural networkcorresponds to applying single-qubit rotations, with the rotation angles beingdependent on the weights and measurement outcomes of the previous layer. Thisrealization has the advantage of being smoothly tunable from the purelyclassical limit with no quantum uncertainty (thereby reproducing the classicalneural network exactly) to a quantum case, where superpositions introduce anintrinsic uncertainty in the network. We benchmark this architecture on asubset of the standard MNIST dataset and find a regime of "quantum advantage,"where the validation error rate in the quantum realization is smaller than thatin the classical model. We also consider another approach where quantumness isintroduced via weak measurements of ancilla qubits entangled with the neuronqubits. This quantum neural network also allows for smooth tuning of the degreeof quantumness by controlling an entanglement angle, $g$, with $g=\frac\pi 2$replicating the classical regime. We find that validation error is alsominimized within the quantum regime in this approach. We also observe a quantumtransition, with sharp loss of the quantum network's ability to learn at acritical point $g_c$. The proposed quantum neural networks are readilyrealizable in present-day quantum computers on commercial datasets.