Vision Explainibility
Links Useful
Flow
- DeconvNet (2013)
- Deep Inside Convolutional Networks (2014)
- Guided BackProp (2015) Aka All Conv net
- Building up on Deep Inside Convolutional Networks and DeconvNet
- Salience Map
- Not class discriminative
- Noise
- Not appealing
- CAM
- less noisy
- not class discriminative
- Worked only a restricted set of CNN templates
- Grad-CAM
- class discriminative
- not high res
- Works for any arbitrary CNN
- Occlusion Map
- Same as the next but not very fast
- Guided GradCAM
- DeepLIFT
- [Noise Tunnel](Noise Tunnel.md)
- Smooth-Grad
- [SmoothGrad Square](SmoothGrad Square.md)
- VarGrad
- [Integrated Gradients](Integrated Gradients.md)
- [Proxy Attention](Proxy Attention.md)
- Conductance
Disadvantages
- [The Unreliability of Saliency Methods](The Unreliability of Saliency Methods.md)
- [Interpretation of Neural networks is fragile](Interpretation of Neural networks is fragile.md)
- Fine grained data