Image regression: How to visualize the feature importance of an image in convolutional neural networks

To investigate trained networks, we can use visualization techniques such as Grad-CAM, occlusion sensitivity, LIME, and deep dream.

But the input of these functions require to be string, char, categorical, cell.

For example, map = occlusionSensitivity(net,img,Y), where Y was the predicted value for img, however, this function shows error: Expected input number 3 to be one of these types: string, char, categorical, cell.

Could anyone tell me how to use these functions for regression analysis?

Answer 1

[Aditya Patil]

the third parameter for occlusionSensitivity is label which was predicted for the model. For example,

label = classify(net,X);
scoreMap = occlusionSensitivity(net,X,label);

See the occlusionSensitivity doc page for more details.