Composing graphical models with neural networks · — more flexible models can require slow...

22
Composing graphical models with neural networks (most slides stolen from Matthew James Johnson) 1

Transcript of Composing graphical models with neural networks · — more flexible models can require slow...

Page 1: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Composing graphical models with neural networks

(most slides stolen from Matthew James Johnson)

1

Page 2: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Outline● Motivation and Examples● Stochastic Variational Inference● Structured Variational Auto-Encoders

2

Page 3: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Example: Generative Clustering

DataGMM

3

Page 4: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Example: Generative Clustering

4

Page 5: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

5

Page 6: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Application: Parse mouse behaviour

Aim: Generative model for different mouse behaviour

6

Page 7: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Application: Parse mouse behaviour

Input: Videos

Aim: Come up with a generative model for the behaviour and the video frames

7

Page 8: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Example: Switching Linear Dynamical System

8

Page 9: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Example: Switching Linear Dynamical System

9

Page 10: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Special Cases

GMM LDS

10

Page 11: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

11

Page 12: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Stochastic Variational Inference

12

Page 13: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Mean Field Approximation

Objective function

Equivalent to the Kalman Filter

13

Page 14: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

14

Algorithm

Page 15: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Structured Variational Auto Encoder

15

Page 16: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Mean Field Approximation

Objective function

16

Page 17: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

Variational Auto-Encoder

17

Page 18: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

18

Page 19: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

19

Algorithm

Page 20: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

20

Page 21: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

21

Page 22: Composing graphical models with neural networks · — more flexible models can require slow top-down inference — limited inference queries — data- and compute-hungry + recognition

22

Conclusion

Possible applicationsReinforcement Learning and BanditsSemi-supervised Learning


Exploring Network Inference Models

Exploring Network Inference Models

Inference in Threshold Models - maxwell.syr.edu

Inference in Threshold Models - maxwell.syr.edu

Composing Feature Models

Composing Feature Models

Diatom-based TP and TN inference models and based TP and TN Inference Models and... · UNCORRECTED PROOF Diatom-based TP and TN inference models and indices for monitoring nutrient

Diatom-based TP and TN inference models and based TP and TN Inference Models and... · UNCORRECTED PROOF Diatom-based TP and TN inference models and indices for monitoring nutrient

Composing Models with Six Different Tools: a Comparative Study

Composing Models with Six Different Tools: a Comparative Study

Demo Light for Composing Models - staff.fnwi.uva.nl · Demo Light for Composing Models Jan van Eijck with contributions by Lakshmanan Kuppusamy and Floor Sietsma May 17, 2011

Demo Light for Composing Models - staff.fnwi.uva.nl · Demo Light for Composing Models Jan van Eijck with contributions by Lakshmanan Kuppusamy and Floor Sietsma May 17, 2011

models for composing a minuet.eckert.pdf

models for composing a minuet.eckert.pdf

Models of collective inference

Models of collective inference

Composing analysis patterns to build complex models: Flight

Composing analysis patterns to build complex models: Flight

Composing graphical models with neural networks …mlg.postech.ac.kr/~readinglist/slides/20161101.pdf2016/11/01  · Composing graphical models with neural networks for structured

Composing graphical models with neural networks …mlg.postech.ac.kr/~readinglist/slides/20161101.pdf2016/11/01  · Composing graphical models with neural networks for structured

Fuzzy Inference Systems. Content The Architecture of Fuzzy Inference Systems Fuzzy Models: – Mamdani Fuzzy models – Sugeno Fuzzy Models – Tsukamoto Fuzzy.

Fuzzy Inference Systems. Content The Architecture of Fuzzy Inference Systems Fuzzy Models: – Mamdani Fuzzy models – Sugeno Fuzzy Models – Tsukamoto Fuzzy.

Embedding population dynamics models in inference

Embedding population dynamics models in inference

Graphical Probability Models for Inference and Decision …mason.gmu.edu/~klaskey/.../GraphicalModels_Unit4_JTInference.pdf · Graphical Probability Models for Inference and Decision

Graphical Probability Models for Inference and Decision …mason.gmu.edu/~klaskey/.../GraphicalModels_Unit4_JTInference.pdf · Graphical Probability Models for Inference and Decision

Memory Bounded Inference on Topic Models

Memory Bounded Inference on Topic Models

Inference for misspecified models - Richard Wilkinson · 2021. 1. 24. · Inference for misspeci ed models Richard Wilkinson University of She eld. Mechanistic models Models describe

Inference for misspecified models - Richard Wilkinson · 2021. 1. 24. · Inference for misspeci ed models Richard Wilkinson University of She eld. Mechanistic models Models describe

Graphical Probability Models for Inference and …mason.gmu.edu/~klaskey/GraphicalModels/GraphicalModels_Unit3_KRep.pdfGraphical Probability Models for Inference and Decision Making

Graphical Probability Models for Inference and …mason.gmu.edu/~klaskey/GraphicalModels/GraphicalModels_Unit3_KRep.pdfGraphical Probability Models for Inference and Decision Making

Bayesian Inference, Generative Models, and …web.stanford.edu/.../BayesianInfGenModsProbCompsNNets.pdfBayesian Inference, Generative Models, and Probabilistic Computations in Interactive

Bayesian Inference, Generative Models, and …web.stanford.edu/.../BayesianInfGenModsProbCompsNNets.pdfBayesian Inference, Generative Models, and Probabilistic Computations in Interactive

Gaussian Models: Bayesian Inference

Gaussian Models: Bayesian Inference

Preface MODELS, METHODS, AND MEASUREMENT · 2020-06-08 · MODELS, METHODS, AND MEASUREMENT 136 EXTRASTATISTICAL INFERENCE Extrastatistical inference, and statistical inference, are

Preface MODELS, METHODS, AND MEASUREMENT · 2020-06-08 · MODELS, METHODS, AND MEASUREMENT 136 EXTRASTATISTICAL INFERENCE Extrastatistical inference, and statistical inference, are

David Duvenaud arXiv:1603.06277v2 [stat.ML] 24 May 2016 · 2018-11-09 · Composing graphical models with neural networks for structured representations and fast inference Matthew

David Duvenaud arXiv:1603.06277v2 [stat.ML] 24 May 2016 · 2018-11-09 · Composing graphical models with neural networks for structured representations and fast inference Matthew