According to one embodiment, a learning device includes an acquirer, a generator, a first identifier, an extractor, a second identifier, and a learning-processor. The acquirer acquires real data. The generator generates pseudo data of the real data using a first neural network. The first identifier identifies whether input data is the real data or the pseudo data. The extractor extracts features of data from the input data. The second identifier identifies whether the features extracted by the extractor are features of the real data or features of the pseudo data. The learning-processor learns the first neural network such that the pseudo data and the real data are not discriminated on the basis of an identification result of each identifier.