Measuring the perceptual availability of phonological features during language acquisition using unsupervised binary stochastic autoencoders