activation atlas: showing all the lower dimension (human understandable) representation of the features learned and how the model organises them
embedding space: provide that the vector (direction and distance) between features have semantic meaning. (King,Queen & Man, Woman - have the same distance)
the first layer has multiple kernels, and 3 layers (RGB) in the input image. now the second layer has same multiple layers which become dimensions for the third layer. (so 96 kernels translate to 96 dimensions for the next layer)