tldrafael / CNNVisualizations

Plain implementation of Grad-CAM, gradient ascent, and adversarial examples for semantic segmentation and classification.

README

See the repo tutorial on .

The topics approached on this repo are:

• Region importance at any layer using Grad-CAM.
  • With or without guided-backpropagation.
• Optimize an input to maximize a neuron or an output with gradient ascent.
  • Deep dream, optimization starting from an existing image.
• Generate adversarial examples.

The results are presented for classification and semantic segmentation.

Semantic Segmentation

The images bellow illustrate the RTK Dataset.

GradCAM

Input Synthesis

Adversarial Examples

The rows follows the respective order: the adversarial example, the isolated noised (scaled 3x for better visualization), and the prediction results.

Classification

The images bellow illustrate the ImageNet.

GradCAM

Input Synthesis

Adversarial Examples

• The noise was scaled 10x for better visualization. The adversarial predictions were: 'malinois', 'screw', 'matchstick', 'dial telephone', and 'briard'.

References

• https://github.com/jacobgil/pytorch-grad-cam/
• https://aman.ai/cs231n/visualization/
• http://cs231n.stanford.edu/slides/2021/lecture_14.pdf
• https://distill.pub/2017/feature-visualization/