Uploaded March 2019 | Updated September 2026, 1 week ago
slides: mlvu.github.io/lectures/51.Deep%20Learning2.annotated.pdf
course materials: mlvu.github.io
Today we discuss neural networks that can generate things that are similar to the data they were fed: generative models. We start out with a discussion of how to define a generator neural network, and why we can't train them in a straightforward way. Then we discuss two training approaches: generative adversarial networks and (variational) autoencoders.
For the variational autoencoder, we derive the whole loss function from first principles, showing how it approximates a maximum likelihood fit on the data.
slides: mlvu.github.io/lectures/51.Deep%20Learning2.annotated.pdf
course materials: mlvu.github.io
Today we discuss neural networks that can generate things that are similar to the data they were fed: generative models. We start out with a discussion of how to define a generator neural network, and why we can't train them in a straightforward way. Then we discuss two training approaches: generative adversarial networks and (variational) autoencoders.
For the variational autoencoder, we derive the whole loss function from first principles, showing how it approximates a maximum likelihood fit on the data.
