14
Learning the structure of Deep sparse Graphical Model Ryan Prescott Adams Hanna M Wallach Zoubin Ghahramani Presented by Zhengming Xing Some pictures are direct ly copied from the paper and Hanna Wallach’s slid es

Learning the structure of Deep sparse Graphical Model

  • Upload
    adin

  • View
    36

  • Download
    0

Embed Size (px)

DESCRIPTION

Learning the structure of Deep sparse Graphical Model. Ryan Prescott Adams Hanna M Wallach Zoubin Ghahramani. Presented by Zhengming Xing. Some pictures are directly copied from the paper and Hanna Wallach’s slides. outline. Introduction Finite belief network Infinite belief network - PowerPoint PPT Presentation

Citation preview

Page 1: Learning the structure of Deep sparse Graphical Model

Learning the structure of Deep sparse Graphical Model

Ryan Prescott Adams Hanna M Wallach

Zoubin Ghahramani

Presented by Zhengming Xing

Some pictures are directly copied from the paper and Hanna Wallach’s slides

Page 2: Learning the structure of Deep sparse Graphical Model

outline

• Introduction

• Finite belief network

• Infinite belief network

• Inference

• Experiment

Page 3: Learning the structure of Deep sparse Graphical Model

Introduction Main contribution: combine deep belief network and nonparametric bayesian together.

Main idea: use IBP to learn the structure of the network

Structure of the network include:

Depth

Width

Connectivity

Page 4: Learning the structure of Deep sparse Graphical Model

Single layer networkUse Binary matrix to represent the network.

Black refer to 1(two unit were connected)

White refer to 0 (two unit were not connected)

IBP can be used as the prior for infinite columns binary matrix

Z

Page 5: Learning the structure of Deep sparse Graphical Model

Review IBP

)(poisson

)1/( nnk

1.First customer tries dishes.

2. Nth customer tries

Tasked dishes K with probability

new dishes))1/(( npoisson

1/ nnk

Page 6: Learning the structure of Deep sparse Graphical Model

Multi-layer network

Page 7: Learning the structure of Deep sparse Graphical Model

Cascading IBP),( Also parameterize by

Each dishes in the restaurant is also a customer in another Indian buffet process

Each matrix is exchangeable both rows and columns

This chain can reach the state with probability one ( number of unit in layer m)

Properties:

For unit in layer m+1

Expected number of parents:

Expected number of children:

0)( mK

K

k kK

1 1/

)(mK

Page 8: Learning the structure of Deep sparse Graphical Model

Sample from the CIBP prior

Page 9: Learning the structure of Deep sparse Graphical Model

model)()1(11 )( mmmmm ZWy m refer to the layers and increase upto M.

1)1)/(exp(2(.)

),0(~)( )()()()()(

x

Ny mk

mk

mk

mk

mk

weights bias

Place layer wise Gaussian prior on weights and bias, Gamma prior on noise precision

Page 10: Learning the structure of Deep sparse Graphical Model

Inference

Weights, bias, noise variance can be sampled with Gibbs sampler.

Page 11: Learning the structure of Deep sparse Graphical Model

Inference( sample Z)Two step:

1.

2.

Sample existing dishes

MH-sample

Add a new unit and, and insert connection to this unit with

For a exist unit remove the connection to this unit with

MH ratio

MH ratio

Page 12: Learning the structure of Deep sparse Graphical Model

Experiment result

Olivetti faces

Remove bottom halves of the test image.

Page 13: Learning the structure of Deep sparse Graphical Model

Experiment result

MNIST Digits

Page 14: Learning the structure of Deep sparse Graphical Model

Experiment resultFrey Faces