The Origin of Grad-CAM AI Study Meeting #4 @Eaglys on 2020/10/25 Shintaro Yoshida @sht_47

The Features of Grad-CAM ● Grad-CAM(Gradient-weighted Class Activation Mapping, 2016, Ramprasaath) ○ Most Famous Method in XAI ( I described the reason in later slide) ○ Update CAM(2015, Zhou) 、Generalize to Any Kind of CNN Architecture ● The Goal of XAI(Explainable Artificial Intelligence) Identify the Mode of Failure (AI << Human) Predict with more Confidence (AI ≒ Human) AI teaches Human (AI >> Human)

The Content - The referred Paper of Grad-CAM - - - Grad-CAMのモデル中身 - Result and Discussion - Implement with Pytorch and Google Colaboratory

NIN(Network In Network, 2014 Lin et al) - Proficient Paper because of two great ideas Introduce 1x1 Conv to reduce the calculation cost ( Applied to InceptionNet、ResNet Botttleneck Block) Introduce GAP(Global Average Pooling) → Recently Adaptive Average Pooling is used ● GAP Performed as a Structural Regularizer ○ More Native to the correspondence between Feature Map and Category ○ NO Added Parameter ○ Robust to Spatial Translation

Object Detectors Emerge In Deep Scene Cnns(2015 Zhou et al) - CNN Model Scene Recognition → Object Detector Emerges No Supervised Dataset of Object Classification and Detection In Previous Research, Object Classification → Object Localization Places Database (2014 Zhou et al )

CAM(Class Activation Mapping 2015 Zhou et al) … … Final Conv GAP FC K Featuer Maps K Element … C class a a 1 Generate CAM Using

CAM(Class Activation Mapping) … … Final Conv GAP FC 4096 Feature Maps 4096 Element … 1000 Class VGG16 (ImageNet) 7 7

Math Equation and Concept of CAM Sum with i, j Weighted Sum with k Each Process is Independent Z is size of Feature Map (Z=49)

Usage of CAM( After Inference) Average With i, j (Image Source : Zhou et al 2015) CAM Weighted Sum with k Inference Generate CAM Weighted Sum with k

Guided Back-Propagation(2015 Springenberg) - Deconvolutional Network (2011 Zeiler) Opposite Process of Max Pooling - Guided Backprop Combine with DeconvNet and ReLU BackPropagation

Result of Guided-Backprop Batch Size : 64 Learning Rate : 0.01 Weight Decay : 0.001 Optimizer : SGD Conv6 Conv9

Grad-CAM(2016 Ramprasaath) CAM limits with GAP → Grad-CAM generalize to Any Architecture Combine CAM(Corase) with Guided-Backprop(Fined-Grained) Insert ReLU to CAM(Only Positive Value is enough) No need to Architectural Change and Re-Train Sum with i and j Weighted Sum with Weighted Sum with

Result 1 of Grad-CAM - Microsoft COCO Dataset - Sample from Validation Dataset - Mistake with Ice Cream

Result 2 of Grad-CAM Mistake at VGG@ImageNet Whether the model has bias or not

Implement - Pytorch 1.6 https://github.com/sht47/grad-cam-Pytorch1.6 - Tensorflow 2.3 (Under Construction) https://github.com/sht47/grad-cam-Tensorflow2.3