Computational and Experimental Methods

Identity-Based Patterns in Deep Convolutional Networks: Generative Adversarial Phonology and Reduplication

Gašper Beguš

This paper models unsupervised learning of an identity-based pattern (or copying) in speech called reduplication from raw continuous data with deep convolutional neural networks. We use the ciwGAN architecture (Beguš, 2021a) in which learning of meaningful representations in speech emerges from a requirement that the CNNs generate informative data. We propose a technique to wug-test CNNs trained on speech and, based on four generative tests, argue that the network learns to represent an identity-based pattern in its latent space. By manipulating only two...

Generative Adversarial Phonology: Modeling Unsupervised Phonetic and Phonological Learning With Neural Networks

Gašper Beguš

Training deep neural networks on well-understood dependencies in speech data can provide new insights into how they learn internal representations. This paper argues that acquisition of speech can be modeled as a dependency between random space and generated speech data in the Generative Adversarial Network architecture and proposes a methodology to uncover the network's internal representations that correspond to phonetic and phonological properties. The Generative Adversarial architecture is uniquely appropriate for modeling phonetic and phonological learning because the network is...

CiwGAN and fiwGAN: Encoding information in acoustic data to model lexical learning with Generative Adversarial Networks

Gašper Beguš

How can deep neural networks encode information that corresponds to words in human speech into raw acoustic data? This paper proposes two neural network architectures for modeling unsupervised lexical learning from raw acoustic inputs: ciwGAN (Categorical InfoWaveGAN) and fiwGAN (Featural InfoWaveGAN). These combine Deep Convolutional GAN architecture for audio data (...

Beguš speaks at USC

October 5, 2021

Gašper Beguš gave a colloquium talk at USC Linguistics on October 4 entitled "Deep Learning and Phonology: Comparing Behavioral and Neural Speech Data with Outputs of Deep Generative Models."

Beguš publishes in Computer Speech & Language

September 16, 2021

Congrats to Gašper Beguš on the publication of his article "Local and non-local dependency learning and emergence of rule-like representations in speech data by deep convolutional generative adversarial networks" in Computer Speech & Language! Click here to download the article (Open Access).

Bleaman publishes in American Speech

May 5, 2021

Congrats to Isaac Bleaman and Dan Duncan (Newcastle University) on the publication of their article "The Gettysburg Corpus: Testing the proposition that all tense /æ/s are created equal" in American Speech. Read it here!

Beguš publishes in Neural Networks

April 21, 2021

Congrats to Gašper Beguš on the publication of his article "CiwGAN and fiwGAN: Encoding information in acoustic data to model lexical learning with Generative Adversarial Networks" in Neural Networks! Click here to download the article (Open Access).

Beguš speaks at ICBS

February 23, 2021

Gašper Beguš will be giving a seminar talk at UC Berkeley's Institute of Cognitive and Brain Sciences on Friday, March 5, from 11:10am to 12pm. The title of his talk is "Modeling Language with Generative Adversarial Networks" and the abstract is below. Click here for more details. Congrats, Gašper!

Can we build models of language acquisition from raw acoustic data in an unsupervised manner? Can deep convolutional neural networks learn to generate speech using linguistically meaningful representations? In this talk, I will argue that language acquisition can be modeled with Generative Adversarial Networks (GANs) and that such modeling has implications both for the understanding of language acquisition and for the understanding of how neural networks learn internal representations. I propose a technique that allows us to wug-test neural networks trained on raw speech. I further propose an extension of the GAN architecture in which learning of meaningful linguistic units emerges from a requirement that the networks output informative data. With this model, we can test what the networks can and cannot learn, how their biases match human learning biases (by comparing behavioral data with networks’ outputs), how they represent linguistic structure internally, and what GAN's innovative outputs can teach us about productivity in human language. This talk also makes a more general case for probing deep neural networks with raw speech data, as dependencies in speech are often better understood than those in the visual domain and because behavioral data on speech acquisition are relatively easily accessible.

Beguš speaks at UC Davis PhonLab

November 4, 2020

Gašper Beguš will be speaking at the UC Davis PhonLab on Friday, Nov 6 at 10AM on the topic "Encoding linguistic meaning into raw audio data with deep neural networks."