Abstract

The Grad-CAM (Gradient-weighted Class Activation Mapping) is a generalization of CAM and is applicable to a significantly broader range of CNN model families.
The intuition is to expect the last convolutional layers to have the best compromise between high-level semantics and detailed spatial information which is lost in fully-connected layers. The neurons in these layers look for semantic class-specific information in the image.

$$L_{Grad-CAM}^c = ReLU(\sum_k\alpha_k^cA^k)$$

where $$\alpha_k^c = \frac{1}{Z}\sum_i\sum_j\frac{\partial{y_c}}{\partial{A_{ij}^k}}$$

Supervised or unsupervised?

• Unsupervised task: generative modeling is an unsupervised task where the model is not told what kind of patterns to look for in the data and there is no error metric to improve the model.
• Supervised classifier/loss func: the training process of the GAN is posed as a supervised learning problem with the help of a discriminator.

Discrimitive | Generative | Latent Models