Feature or Activation Maps

Feature or activation map is the non-linear output of a convolution layer. It’s the resulting matrix of the convolution operation, where we see what have stimulated the layer’s neurons.

On CLIP code, “logits” means the feature map, with non-normalized features, which are then passed to softmax in order to map each class to one output value.

Refs

• What is the meaning of the word logits in TensorFlow?
• What is the definition of a "feature map" (aka "activation map") in a convolutional neural network?