Similarity comparison: sample vs hidden-layer activation

Draw a digit and submit. The left row shows training cases retrieved by raw-pixel (feature) similarity; the right row shows cases retrieved by the neural network's hidden-layer activation similarity. Feature similarity reflects how the images look; activation similarity reflects how the model represents them internally. Green outlines mark neighbors whose label matches the prediction.

Centered input:
Sample similarity (raw pixels / feature space)
Activation similarity (model representation)