Latent variable generative models (LVGM)
from https://acdl2019.icas.xyz/phillip-isolas-lectures/
Interests
- LVGM as model of brain
- hierarchical generative models (neuroscience motivation)
- constraints and inductive biases needed for human-like representations
- disentangled
- unsupervised vs task optimised
- compression with generative models
- relating recognition models to neural data
- LVGM as tool for modeling measurements
- learn low dimensional manifolds on neural data
- hope for interpretable latents
Variational Autoencoders
from https://acdl2019.icas.xyz/phillip-isolas-lectures/
Why do we find VAEs interesting?
- approximate generative models for naturalistic data
- approximate inference over latent variables
- principled motivation for approximations
- representations in some simple cases seem to be what we hoped for