Composing graphical models with neural networks
(most slides stolen from Matthew James Johnson)
1
Outline● Motivation and Examples● Stochastic Variational Inference● Structured Variational Auto-Encoders
2
Example: Generative Clustering
DataGMM
3
Example: Generative Clustering
4
5
Application: Parse mouse behaviour
Aim: Generative model for different mouse behaviour
6
Application: Parse mouse behaviour
Input: Videos
Aim: Come up with a generative model for the behaviour and the video frames
7
Example: Switching Linear Dynamical System
8
Example: Switching Linear Dynamical System
9
Special Cases
GMM LDS
10
11
Stochastic Variational Inference
12
Mean Field Approximation
Objective function
Equivalent to the Kalman Filter
13
14
Algorithm
Structured Variational Auto Encoder
15
Mean Field Approximation
Objective function
16
Variational Auto-Encoder
17
18
19
Algorithm
20
21
22
Conclusion
Possible applicationsReinforcement Learning and BanditsSemi-supervised Learning