Top-down

Neural Attention

by Excitation Backprop

We propose Excitation Backprop, a backpropagation-based method to visualize a CNN's top-down attention. Our method is based on a probabilistic Winner-Take-All formulation, which is inspired by the Selective Tuning attention model.

Paper

[PDF] [Supplementary][arXiv]

Jianming Zhang, Zhe Lin, Jonathan Brandt, Xiaohui Shen and Stan Sclaroff. "Top-down Neural Attention by Excitation Backprop." To appear in Proc. European Conference on Computer Vision, 2016.

(Accepted for oral presentation)

 

[BibTex]

@inproceedings{zhang2016EB,

  title={Top-down Neural Attention by Excitation Backprop},

  author={Zhang, Jianming and Lin, Zhe and Brandt, Jonathan, Shen, Xiaohui and Sclaroff, Stan},

  booktitle={European Conference on Computer Vision(ECCV)},

  year={2016},

}

Contact: jmzhang AT bu.edu

Code

We provide a Caffe implementation of the Excitation Backprop method. If you use this code, please cite our paper.

Notice: This software implementation is provided for academic research and non-commercial purposes only.  This implementation is provided without warranty.  The Excitation Backprop method described in the above paper and implemented in this software is patent-pending by Adobe.

Code: [Github]

Slides