: Discard the final fully connected layer of the network. Instead of a single "spider" label, you want the activation values from the last pooling layer.

: Behaviors like constructing decoys out of debris, which create distinct visual signatures.

: Input your images from the .rar file into the network. The resulting output vector (often 512, 1024, or 2048 dimensions) is your "deep feature."

When analyzing spider imagery, your deep features should ideally capture: