Explaining the Unexplained: A CL CLass-Enhanced Attentive Response - - PowerPoint PPT Presentation

▶

Sep 13, 2023 109 likes •279 views

Explaining the Unexplained: A CL CLass-Enhanced Attentive Response (CLEAR) Approach to Understanding Deep Neural Networks Devinder Kumar* , Alexander Wong & Graham W. Taylor Current Approaches Heatmap/Attention based! Deconvolution:

SLIDE 1

Explaining the Unexplained: A CL CLass-Enhanced Attentive Response (CLEAR) Approach to Understanding Deep Neural Networks

Devinder Kumar*, Alexander Wong & Graham W. Taylor

SLIDE 2

Current Approaches – Heatmap/Attention based!

Saliency: Simonyan et.al. CVPR 2013 Guided backpropagation: ICLR 2015 Deconvolution: Zeiler et.al. ECCV 14 Deep Taylor Decomp. Montavon et. al. PR journal 2017 Prediction Difference: Zintgraf et. al. ICLR 2017

SLIDE 3

Input Output

3

Heatmap

DNN

2 3

Focuses on right areas: Looks correct!

Focuses on wrong part, curve might be two; but why not 3 or 5 or 6? Probably focuses

n correct part,

but why 3?

Interpretation

DNN DNN

Interpretations

SLIDE 4

Binary Heatmap CLEAR Map

Class Enhanced Attentive Response (CLEAR) Map

SLIDE 5

given Output response at layer l Output response of layer l Dominant Class Response Dominant Response

Class Enhanced Attentive Response (CLEAR)

SLIDE 6

Class Enhanced Attentive Response (CLEAR) Map

SLIDE 7

Input Output

Heatmap

DNN

2 3

Focuses on right areas: Looks correct! Focuses on wrong part, curve might be two; but why not 3 or 5 or 6? Probably focuses

n correct part,

but why 3?

CLEAR map

Major part of the positive focus is of 3 Major part of the positive focus represents 2 Major part of negative focus is 3; higher activation than any other class

Interpretation Interpretation

CLEAR color map

1 2 3 4 5 6 7 8 9

DNN DNN

SLIDE 8

MNIST RESULTS

Correctly Classified Wrongly Classified

SLIDE 9

SVHN RESULTS

Correctly Classified Wrongly Classified

SLIDE 10

MNIST & SVHN RESULTS

SLIDE 11

Stanford Dog Dataset Results

SLIDE 12

Stanford Dog Dataset Results

SLIDE 13

Stanford Dog Dataset Results

SLIDE 14

Conclusion

Sparsity in the individual response maps from the last layer kernels :

same pattern for all datasets considered.

Evidence for classes tend to come from very specific localized regions.
CLEAR maps enable the visualization of not only the areas of interest

that predominantly influence the decision-making process, but also the degree of influence as well as the dominant class of influence in these areas.

Showed efficacy of CLEAR maps both quantitatively and qualitatively.

SLIDE 15

Explaining the Unexplained: A CL CLass-Enhanced Attentive Response - - PowerPoint PPT Presentation

Explaining the Unexplained: A CL CLass-Enhanced Attentive Response (CLEAR) Approach to Understanding Deep Neural Networks

Devinder Kumar*, Alexander Wong & Graham W. Taylor

Current Approaches – Heatmap/Attention based!

Input Output

3

Heatmap

2 3

Interpretation

Interpretations

Binary Heatmap CLEAR Map

Class Enhanced Attentive Response (CLEAR) Map

Class Enhanced Attentive Response (CLEAR)

Class Enhanced Attentive Response (CLEAR) Map

MNIST RESULTS

SVHN RESULTS

MNIST & SVHN RESULTS

Stanford Dog Dataset Results

Stanford Dog Dataset Results

Stanford Dog Dataset Results

Conclusion

same pattern for all datasets considered.

that predominantly influence the decision-making process, but also the degree of influence as well as the dominant class of influence in these areas.

Thank You!

devinder.kumar@uwaterloo.ca http://devinderkumar.com