Convolutional Poisson Gamma Belief Network Chaojie Wang Bo Chen - - PowerPoint PPT Presentation

convolutional poisson gamma belief network
SMART_READER_LITE
LIVE PREVIEW

Convolutional Poisson Gamma Belief Network Chaojie Wang Bo Chen - - PowerPoint PPT Presentation

Convolutional Poisson Gamma Belief Network Chaojie Wang Bo Chen Sucheng Xiao Mingyuan Zhou National Laboratory of Radar Signal Processing, Xidian University, Xi'an, Shaanxi, China McCombs School of Business, The


slide-1
SLIDE 1

National Lab of Radar Signal Processing

♖ National Laboratory of Radar Signal Processing, Xidian University, Xi'an, Shaanxi, China ♘ McCombs School of Business, The University of Texas at Austin, Austin, TX, USA

Chaojie Wang♖ Bo Chen♖ Sucheng Xiao♖ Mingyuan Zhou♘

1

Convolutional Poisson Gamma Belief Network

2019-6-11

Xidian University & UT-Austin

slide-2
SLIDE 2

National Lab of Radar Signal Processing

ü Preserve all textual information ※ Extremely large and sparse matrices ※ Burdens of calculation and storage ※ Difficult to model directly ü Term-document frequency count matrix ※ Lose word order ü Project words to low-dimensional vectors ※ Require additional large corpora

q Simplified Lossy Representation

Simplified

Motivation

“I love it” Document One-hot Sequence don't hate I it love I love it

1                 1                 1                

Most basic representation Ø A sequence of one-hot vectors Ø Bag-of-words Ø Word embeddings

Document Representation

q Basic Lossless Representation

Challenge 2019-6-11

2

Xidian University & UT-Austin

slide-3
SLIDE 3

National Lab of Radar Signal Processing

Our Contribution

Convolutional Poisson Factor Analysis

q Generative model of CPFA

ü Preserve word order information ü Directly model sparse matrices ü Take advantages of the sparsity ü Support parallel computation ü Capture pharse-level topics

Probabilistic Convolutional Layer

2019-6-11

3

Xidian University & UT-Austin

slide-4
SLIDE 4

National Lab of Radar Signal Processing

Our Contribution

Convolutional Poisson Gamma Belief Network

q Generative model of CPGBN q Probabilistic Pooling Layer

ü Transfer the messages from deeper layers ü Jointly Train all the other layers ü Deep extention can boost performance ü Hierachical pharse-level topic Equivalent

2019-6-11

4

Xidian University & UT-Austin

slide-5
SLIDE 5

National Lab of Radar Signal Processing

Our Contribution

Hybrid MCMC/Variational Inference

q Convolutional inference network

ü Fast in out-of-sample prediction ü Parallel scalable inference ü Easy extension (e.g., modeling document labels)

q Weibull Reparameterization

Approximate

2019-6-11

5

Xidian University & UT-Austin

slide-6
SLIDE 6

National Lab of Radar Signal Processing

Experiment

Phrase-level Topics Visualization

2019-6-11

6

Xidian University & UT-Austin

slide-7
SLIDE 7

National Lab of Radar Signal Processing

♖ National Laboratory of Radar Signal Processing, Xidian University, Xi'an, Shaanxi, China ♘ McCombs School of Business, The University of Texas at Austin, Austin, TX, USA

Chaojie Wang♖ Bo Chen♖ Sucheng Xiao♖ Mingyuan Zhou♘

Thank you !

2019-6-11

7

Xidian University & UT-Austin