"Outskirts of Deep Generative Modeling"
Presented By: Colin Raffel, Google Brain
Generative modeling is a branch of machine learning which seeks to uncover the underlying structure in a dataset by training a model to generate new, similar data. Deep generative models leverage the expressive power of neural networks to allow modeling of complex natural data like images, text, and speech. In this talk, I will give an overview of the dominant paradigms for deep generative modeling with an emphasis on my own work on learned representations and sequential data. First, I will cover sequence-to-sequence models, which are trained to generate a transformation of an input sequence. I will show how sequence-to-sequence models can be used to transform their input “online,” as the input is still being generated. Next, I will discuss variational autoencoders, which learn a mapping to a compressed representation that can capture the high-level characteristics of a data point. I will show how variational autoencoders can be applied to generating long musical sequences by using a hierarchical neural network architecture. I will then turn to generative adversarial learning, a framework for learning a similarity measure between two sets of data. I will show how we used this approach to test whether a generative model which provides control over its data-generation process can also produce useful learned representations. Finally, I will close with a look beyond the outskirts of deep generative modeling and argue for a focus on models which can uncover, represent, and modify the latent structure in data.
Colin Raffel is currently a Senior Research Scientist at Google Brain, where he works in Ian Goodfellow's group. He received his Ph.D. from Columbia University in 2016 where he was supervised by Daniel P. W. Ellis. His research is focused on generative modeling and representation learning with an emphasis on sequential data like music, speech, and text. He also works to improve experimental practice by releasing software tools and conducting meta-studies on evaluation.
This is a joint seminar with the Departments of Computer Science and Computer and Mathematical Sciences, UTSC