Last Layer

• L2 Nearest neighbors in feature space (in last layer)

Activations

Visualizing Activations

어떤 patch는 사람의 얼굴 부분에서 활성화됨

Maximally Activating Patches

• Pick a layer and a channel
• Visualize image patches that correspond to maximize activations
• 각 레이어는 이미지의 일부분을 봄
• 깊어질수록 더 복잡한 물체를 인식
• 위(Conv3)의 1번째 줄 → 동그라미 (아마 눈)
• 아래(Conv9)의 2번째 줄 → 사람 얼굴

Saliency

• Salient: 중요한

• Saliency: the quality of being particularly noticeable or important; prominence

• Saliency Map

  

Perturbation Analysis

• Mask part of the image before feeding to CNN, check how much predicted probabilities change

  

  → If probability drop significantly, masked region is important

Guided Backpropagation

• 중요한 부분만을 Gradient Descent

• Images come out nicer if only backprop positive gradients through each ReLU (guided backprop)

  • ex. 개 예측 → negative: 개 X 영역, positive: 개 O 영역

• 계산 예시