[深度学习论文笔记][Neural Arts] Inceptionism: Going Deeper into Neural Networks

Mordvintsev, Alexander, Christopher Olah, and Mike Tyka. “Inceptionism: Going deeper into neural networks.” Google Research Blog. Retrieved June 20 (2015). (Citations: 36).

1 Motivation

Each layer of the network deals with features at a different level of abstraction. We feed the network an arbitrary image and let the network to enhance whatever it detected.

2 Method

Goal



Dream on arbitary layer’s activation A. The gradient wrt A is



3 Results

See Fig. If a cloud looks a little bit like a dog, the network will make it look more like a dog. This in turn will make the network recognize the dog even more strongly on the next
pass and so forth, until a highly detailed dog appears, seemingly out of nowhere.

Lower layers tend to produce strokes or simple ornamentlike patterns, because those layers are sensitive to basic features such as edges and their orientations. Higherlevel layers will identify more sophisticated features in images, complex features or even
whole objects tend to emerge.