Elamur Research
☰
home / papers / 2013 / Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv — non-exclusive license to distribute" does not allow redistributing the text itself. Original — arxiv.org/pdf/1312.6034

Summary

The paper asks how to inspect what a deep ConvNet image classifier has learned. Both proposed techniques rely on the gradient of a class score with respect to the input pixels. The first applies gradient ascent, following Erhan et al., to synthesize an L2-regularised image that maximizes an unnormalised class score, producing a picture of how the network represents that category. The second treats the gradient at a particular image as a local linear approximation of the score and turns its magnitudes into a per-pixel class saliency map, which needs only one backward pass. The experiments use a cuda-convnet model trained on ILSVRC-2013. The saliency maps serve as seeds for GraphCut segmentation, giving weakly supervised object localisation without any box or mask annotations. The authors also show that DeconvNet reconstruction is essentially gradient back-propagation, differing only in how ReLU is handled. This makes gradient-based visualisation a more general tool, since it applies to any layer. The work became a foundational reference for saliency-based explanation of neural networks.

Key points

  • The model was trained on ILSVRC-2013 (1.2M images, 1000 classes) with layers conv64-conv256-conv256-conv256-conv256-full4096-full4096-full1000. Its validation top-1/top-5 error was 39.7%/17.7%, slightly better than the single-net result of Krizhevsky et al. (40.7%/18.2%).
  • Optimising raw class scores instead of soft-max posteriors keeps the procedure from raising one class simply by suppressing the others. For RGB inputs, saliency is taken as the maximum absolute gradient across colour channels, then averaged over 10 crops.
  • For localisation, the foreground colour model (a GMM) is fitted to pixels above the 95% saliency quantile and the background model to pixels below the 30% quantile. After GraphCut, the largest connected foreground component becomes the object mask, and its bounding box is reported.
  • This method reached 46.4% top-5 localisation error in the ILSVRC-2013 challenge. That beat the authors' own fully supervised ILSVRC-2012 entry (50.0%) but fell short of the supervised winner (29.9%).
  • In the DeconvNet comparison, convolution and max-pooling "switches" correspond exactly to gradient computation. Only the ReLU step differs, because DeconvNet gates on the output reconstruction rather than on the layer input.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Other

Cited by

270 papers of our library cite it (of the 140673 whose bibliographies we read); 8 of them have a page here, 262 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1312.6034 · arXiv 1312.6034

@misc{simonyan2013deep,
  title = {{Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps}},
  author = {Simonyan, Karen and Vedaldi, Andrea and Zisserman, Andrew},
  year = {2013},
  doi = {10.48550/arxiv.1312.6034},
  eprint = {1312.6034},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1312.6034}
}

References

13 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. D. Baehrens, T. Schroeter, S. Harmeling, M. Kawanabe, K. Hansen, and K.-R. Müller. How to explain individual classification decisions. JMLR, 11:1803–1831, 2010. find
  2. A. Berg, J. Deng, and L. Fei-Fei. Large scale visual recognition challenge (ILSVRC), 2010. URL http://www.image-net.org/challenges/LSVRC/2010/. ↗
  3. Y. Boykov and M. P. Jolly. Interactive graph cuts for optimal boundary and region segmentation of objects in N-D images. In Proc. ICCV, volume 2, pages 105–112, 2001. find
  4. D. C. Ciresan, U. Meier, and J. Schmidhuber. Multi-column deep neural networks for image classification. In Proc. CVPR, pages 3642–3649, 2012. find
  5. D. Erhan, Y. Bengio, A. Courville, and P. Vincent. Visualizing higher-layer features of a deep network. Technical Report 1341, University of Montreal, Jun 2009. find
  6. P. Felzenszwalb, D. Mcallester, and D. Ramanan. A discriminatively trained, multiscale, deformable part model. In Proc. CVPR, 2008. find
  7. G. E. Hinton, S. Osindero, and Y. W. Teh. A fast learning algorithm for deep belief nets. Neural Computation, 18(7):1527–1554, 2006. find
  8. A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet classification with deep convolutional neural networks. In NIPS, pages 1106–1114, 2012. find
  9. Q. Le, M. Ranzato, R. Monga, M. Devin, K. Chen, G. Corrado, J. Dean, and A. Ng. Building high-level features using large scale unsupervised learning. In Proc. ICML, 2012. find
  10. Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. find
  11. F. Perronnin, J. Sánchez, and T. Mensink. Improving the Fisher kernel for large-scale image classification. In Proc. ECCV, 2010. find
  12. K. Simonyan, A. Vedaldi, and A. Zisserman. Deep Fisher networks and class saliency maps for object classification and localisation. In ILSVRC workshop, 2013. URL http://image-net.org/challenges/LSVRC/2013/slides/ILSVRC\_az.pdf. ↗
  13. M. D. Zeiler and R. Fergus. abs/1311.2901v3, 2013. Visualizing and understanding convolutional networks. CoRR, find our card