Boltzmann Machines constitute a class of neural networks with applications to\nimage reconstruction, pattern classification and unsupervised learning in\ngeneral. Their most common variants, called Restricted Boltzmann Machines\n(RBMs) exhibit a good trade-off between computability on existing silicon-based\nhardware and generality of possible applications.\n Still, the diffusion of RBMs is quite limited, since their training process\nproves to be hard. The advent of commercial Adiabatic Quantum Computers (AQCs)\nraised the expectation that the implementations of RBMs on such quantum devices\ncould increase the training speed with respect to conventional hardware. To\ndate, however, the implementation of RBM networks on AQCs has been limited by\nthe low qubit connectivity when each qubit acts as a node of the neural\nnetwork.\n Here we demonstrate the feasibility of a complete RBM on AQCs, thanks to an\nembedding that associates its nodes to virtual qubits, thus outperforming\nprevious implementations based on incomplete graphs.\n Moreover, to accelerate the learning, we implement a semantic quantum search\nwhich, contrary to previous proposals, takes the input data as initial boundary\nconditions to start each learning step of the RBM, thanks to a reverse\nannealing schedule. Such an approach, unlike the more conventional forward\nannealing schedule, allows sampling configurations in a meaningful neighborhood\nof the training data, mimicking the behavior of the classical Gibbs sampling\nalgorithm.\n We show that the learning based on reverse annealing quickly raises the\nsampling probability of a meaningful subset of the set of the configurations.\nEven without a proper optimization of the annealing schedule, the RBM\nsemantically trained by reverse annealing achieves better scores on\nreconstruction tasks.\n