24
Visualizing and Understanding Convolution Networks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some sides are modified based on Hamid Izadinia’s slides on Vision seminar on Autumn 2014 1

Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

  • Upload
    others

  • View
    4

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Visualizing and Understanding

Convolution Networks

Authors: Matthew D. Zeiler, Rob Fergus New York University

Presenter: Jason Ren Some sides are modified based on Hamid Izadinia’s slides

on Vision seminar on Autumn 2014

1

Page 2: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Main Contributions

• Give insight into the internal operation &behavior

• Diagnostic & Improve the performance

• Occlusion experiments for spatial understanding

Page 3: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Architecture

Page 4: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Difference with Alex-Net

Small filter size & Small stride #Modified according to Visual Results

Page 5: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

5

Approach

• Interpret the intermediate-layer features activities

• What patch cause activation in feature map

• How?

Page 6: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Deconvnet & Convnet

6

Page 7: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Unpooling

Approximate Inverse

Page 8: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

• Rectification

• Relu

• Filtering

• Transposed Version

8

Page 9: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Feature Visualization

Page 10: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Feature Visualization

Page 11: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Feature Visualization

Page 12: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Feature Visualization

Page 13: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Notes

• Hierarchical representation of features

• Larger invariance in higher layers(Layer 5)

• Selective of discriminative parts of image

Page 14: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Feature Evolution During Training

Page 15: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Feature Invariance

Page 16: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Feature Invariance

Page 17: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Small 1st layer filter & stride

• Layer 1: more coverage of middle frequencies

• Layer 2: less aliasing artifacts

Page 18: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Occlusion Sensitivity

Page 19: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Experiments

Page 20: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Experiments - size

Page 21: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Experiments - generalization

Page 22: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Experiments - generalization

Page 23: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Experiments - feature analysis

Page 24: Visualizingand Understanding ConvolutionNetworksVisualizingand Understanding ConvolutionNetworks Authors: Matthew D. Zeiler, Rob Fergus New York University Presenter: Jason Ren Some

Q & A

Thanks!