When
This coming Friday Sept 22 we welcome Dr. Gasper Begus from UC Berkeley who will be speaking on "Modeling language as a dependency between the latent space and data". This talk is in-person. Dr. Begus is available for meetings on Friday morning and afternoon before the colloquium. Slots are 30 minutes, and up to three people at a time can sign up for one slot. To sign up, please go to:
https://docs.google.com/spreadsheets/d/13vu0MyCUs3ftbUx8KP5K2ahoS1S143ImyiTLxkgLDvs/edit?usp=sharing
Date/Time: Friday Sept 22, 3:00 - 4:30
Location: Communications 311
Modeling language as a dependency between the latent space and data
There are many ways to model language -- with rules, exemplars, finite state automata, or Bayesian approaches. In this talk, I propose a new way to model language in a fully unsupervised way from raw speech: as a dependency between latent space and generated data in generative AI models called Generative Adversarial Nets (GANs). I argue that such modeling has implications both for the understanding of language acquisition and for the understanding of how deep neural networks learn internal representations. I propose an extension of the GAN architecture (fiwGAN) in which meaningful linguistic properties emerge from two networks learning to exchange information. FiwGAN captures the perception-production loop of human speech and, unlike most other deep learning architectures, has traces of communicative intent. I further propose a technique to identify latent variables in deep convolutional networks that represent linguistically meaningful units in a causal, disentangled, and interpretable way. We can thus uncover symbolic-like representations at the phonetic, phonological, syntactic, and lexical semantic levels, analyze how learning biases in GANs match human learning biases in behavioral experiments, how speech processing in the brain compares to intermediate representations in deep neural networks, and what GANs’ innovative outputs can teach us about productivity in human language.