Do androids dream of electric sheep? The answer lies within Google's new image recognition algorithm, DeepDream. While the algorithm is more generally used to identify objects in images, it can also be used to give images a “dreamy” makeover. To fully understand what DeepDream is, and how it gives images these bizarre makeovers, we must dive into the world of “neural network” computer systems that attempt to mimic the problem-solving ability of the human brain. To increase their intuitive ability, neural networks can be trained with study guides and evaluated by tests-- not much different from the way that students learn. Overcoming one of the great challenges of computer science, DeepDream utilizes complex structures to achieve object recognition within provided images.
Introduction
Google recently released DeepDream, an image classification program. Simply put, image classification is the process of analyzing an image and determining what exactly is in it- be it a dog, tree, or balloon. Object recognition is considered to be one of the most difficult challenges facing modern computer scientists. Progress in image classification relies on our ability to truly understand and electronically replicate human sight. DeepDream makes significant strides in improving image classification because of its advanced internal structure that mimics the human brain. One of the more interesting outcomes of DeepDream is an overlay of trippy and bizarre visuals, comparable to drug-induced hallucinations, as it attempts to "see objects", as seen in Fig. 1. But how does image classification lead to these outlandish images? To understand these bizarre image overlays, we must first understand the neural network that constitutes DeepDream’s internal structure.
A. Mordvintsev/Research Blog
Figure 1: Places In Rocks: The above image is produced by a variation of the DeepDream algorithm trained to recognize "places" [1].
