Local and non-local dependency learning and emergence of rule-like representations in speech data by deep convolutional generative adversarial networks

This paper argues that training Generative Adversarial Networks (GANs) on local and non-local dependencies in speech data offers insights into how deep neural networks discretize continuous symbolic-like rule-based morphophonological processes emerge a convolutional architecture. Acquisition of has recently been modeled as dependency between latent space generated by GANs Beguš (2020b), who models learning simple allophonic distribution. We extend this approach to test phonological include approximations morphological processes. further parallel outputs the model results behavioral experiment where human subjects are trained used for GAN network. Four main conclusions emerge: (i) provide useful information computational acquisition even if comparatively small dataset an artificial grammar experiment; (ii) easier learn than processes, which matches both typology world’s languages. also proposes (iii) we can actively observe network’s progress explore effect steps representations keeping constant across different steps. Finally, shows (iv) network learns encode presence prefix with single variable; interpolating variable, operation process. The proposed technique retrieving general implications our understanding suggests rule-like generalizations represented interaction variables space.