Google Deep Dream and the Strong Triadic Closure Property
Google’s Deep Dream is a software that uses neural networks to alter the appearance of an image. The original picture, after being “deep dreamed”, becomes a psychedelic and fantastical rendition bearing only some resemblances to the original. The YouTube video, see sources below, shows examples of the software’s visual output, but I would highly recommend experimenting with the program as it outputs visually stimulating and unique images. Below I will summarize the basic algorithm that creates these images as well as its connection to Cornell University’s Networks course.
The fundamental algorithm of Deep Dream involves using a neural network with a series of connected nodes. Generally speaking, a neural network is trained to evaluate a function on its effectiveness or success. For a good analogy on neural networks see minute 2. In the case of Deep Dream the software looks for the nodes and connections triggered when examining an image—the triggering is indicative of certain visual characteristics such as lines or textures that are most prevalent. The program then intensifies these characteristics over many iterations creating a new image, i.e. creates stronger connections between those nodes. The fundamentals of the process rely on identifying strong and weak connections between nodes, much like The Strong Triadic Closure Property (STCP), which states that if one node is bonded to two other nodes by strong connections, then those two nodes are more likely to be strongly connected. The Deep Dream program works with many more than three nodes, but the underlying principles are analogous with the STCP. With a given network, certain connections can be switched from weak to strong in order to better satisfy the STCP. Deep Dream automatically identifies the specific nodes and connections it wishes to strengthen and adjusts them accordingly.
Deep Dream, with its current application, does not explore situations critical to society; it simply examines the networks describing an images’ underlying appearances. The implications of a system that can autonomously and systematically adjusts a network to increase the prevalence of certain occurrences, however, is far more important. The applications are nearly endless given that human interactions, dynamic systems, and ecosystems can all be represented with networks. Google, is by no means, the only company using neural networks, but their utilization of a neural networks can help pave the way for a more sophisticated understanding and manipulation of networks. For now, enjoy the experience that is Google Deep Dream!
Sources:
http://deepdreamgenerator.com/
http://deepdreamgenerator.com/gallery