Opening Variational Inference, Part 3
Variational Inference, Part 3
- Length
- 18 min
- Level
- Undergraduate
- Recorded
- UC Berkeley2020
Chapter 1 of 3: Amortized variational inference
Amortized variational inference replaces a separate approximate posterior for each data point with one neural network that outputs the parameters of that distribution. Levine, in Berkeley's CS 285 on deep reinforcement learning, asks how to differentiate the variational lower bound with respect to the inference network's parameters. A first answer adapts the policy gradient estimator, which handles continuous and discrete latent variables but has high variance and needs several samples. Levine then derives the reparameterization trick, writing a Gaussian latent sample as a deterministic function of network outputs and independent noise. Backpropagation in PyTorch or TensorFlow then yields low-variance gradients from one sample per data point.

Your comment is kept while you sign up.Sign in
Derives the reparameterization trick for amortized variational inference to reduce gradient estimator variance.
Amortized variational inference replaces a separate approximate posterior for each data point with one neural network that outputs the parameters of that distribution. Levine, in Berkeley's CS 285 on deep reinforcement learning, asks how to differentiate the variational lower bound with respect to the inference network's parameters. A first answer adapts the policy gradient estimator, which handles continuous and discrete latent variables but has high variance and needs several samples. Levine then derives the reparameterization trick, writing a Gaussian latent sample as a deterministic function of network outputs and independent noise. Backpropagation in PyTorch or TensorFlow then yields low-variance gradients from one sample per data point.
This lecture
Lecture 54 of 73 of Deep Reinforcement Learning.
Choose a node to see how it connects to this lecture. What comes before is on the leftat the top. What comes after is on the rightat the bottom. The ring is what it teaches. People and places are above. Other lectures and books are below.
Press a chosen node again to open it.