Grad-CAM: 모델이 보는 곳 확인하기

A CNN can be right for the wrong reason. Grad-CAM reveals which pixels actually drove a prediction, so you can trust or question it.

Grad-CAM paints a heatmap on top of the input. Warm regions are where the model focused; cool regions it ignored. 🔥

It traces the gradient of the chosen class score back to a convolutional layer, measuring how much each feature map mattered.

The best maps come from the last convolutional layer, where features are rich yet still spatial enough to localize.

Average the gradients to get a weight per feature map, then combine the maps with those weights into one coarse map.

A ReLU over the combined map keeps regions that push the class up, discarding pixels that argue against it.

The map is small, so you upsample it to the image resolution before overlaying it for a clear visual.

If the heatmap lights up the background instead of the object, your model learned a shortcut, not the real signal.

Run Grad-CAM on wrong predictions to see what misled the model. The focus region often explains the error instantly.

Clear visual evidence of where a model looks makes it far easier to explain and defend in real deployments.

Grad-CAM is an approximate visualization, not exact math. Treat it as a strong hint, then confirm with other checks. 🪟

Use Grad-CAM to turn gradients into a heatmap over the image, exposing what drove a prediction and catching shortcut learning. 🎯