It removes all the top layers of the network

It disables backpropagation

 It initializes weights randomly

 It loads weights trained on the ImageNet dataset

(None, 32)

(None, 1)

(None, 10)

(None, 224, 224, 32)

 Global feature selection

Local interpretability for a specific prediction

Training accuracy

Data visualization only

ROC curve

Aggregated feature importance

Evolution of prediction from base value

Loss over epochs

Class activation maps

Input image pixels

 First convolutional layer filters

Weights of the dense layer

1024 tokens

256 tokens

512 tokens

 128 tokens

Because softmax increases gradients

Because of softmax

Because 784 input features are too few

Because sigmoid activation causes small gradients in deep layers

Makes model overfit faster

Reduces image size

Hides noise in CT scans

Matches input scale for pre-trained networks

A list of the most common characters in the corpus

A mapping from words to unique integer indices

A mapping from word frequency to sentence length

A mapping from labels to predictions

(None, 224, 224, 3)

(None, 32, 32, 3)

(None, 28, 28, 1)

(None, 64, 64, 3)