USC
About this Article
Written by:
Written on: March 21st, 2017
Tags: art, electrical engineering, computer science, lifestyle, artificial intelligence, computer, google, neural network, neuroscience, brain, dream
Thumbnail by: A. Mordvintsev/Research Blog
About the Author
Also in this Issue
Are You a Human? Exploring What Web Security Means to YouWritten by: Austin Bowie
Engineering Ice CreamWritten by: Stephanie Ego
From Shark Skin to SpeedWritten by: Jane Li
Leaving the Light On: Vacuum Tubes and their Reemergence Written by: Greggory Caine
Stay Connected

Volume XVIII Issue I > Do Androids Dream of Electric Sheep? A Look Into Google's DeepDream
While Fig. 3 may seem strange, it doesn't compare to Fig. 1. In Fig. 1, DeepDream produces an image, and then that same image is fed back into DeepDream. This continuous, “Inception”-like​ looping is what causes the drastic differences between the input and final output [5]. A closer look at Fig. 1 shows the neural network looking for what it's trained to see. After many iterations, the rocks slowly transform into buildings and pagodas. Each iteration of DeepDream morphs the image slightly, as it attempts to "see" the objects it's trained to see. Even if these objects aren't present, as is the case with Fig. 1, the network tries to perceive them from something that looks similar [5]. This feedback loop eventually creates pronounced images which weren't even in the original image. DeepDream and, by extension neural networks, are incredibly powerful. While they have been under active research for over 70 years, they are still in their infancy.

Neural Networks in the Future

While computers routinely fail in some complex areas, it's clear that neural networks show promise in these problem spaces. Consider the pure computing power of computers combined with the plasticity and versatility of the human brain. The possibilities are endless. DeepDream has already shown potential in the field of image recognition and processing. Future advancements could do more than create trippy images - it could lead to robots with human-like vision with artificial eyes [6]. For the more financially inclined, a properly trained neural network could analyze the stock market and predict future trends [6]. In the field of thermodynamics, neural networks could learn to predict vapor-liquid equilibrium data for various substances [7].
The list of potential applications is endless. However, the loftiest goal with neural networks is to advance artificial intelligence [2]. Russell and Norvig believe that neural networks are essential to properly simulating intelligence, a theory which sounds incredibly logical. It intuitively makes sense that the best application of an electronic structure that mimics the brain is to recreate human thought processing. A fully capable artificial brain could introduce a new trend of neural network enhanced androids and ultimately revolutionize the concept of humanity. And who knows- maybe these androids will dream of electric sheep after all! We may be light years away from this concept, but DeepDream is the first step toward understanding the vast capabilities of the neural network.

References

    • [1] A. Mordvintsev, "DeepDream - a code example for visualizing Neural Networks", Research Blog, 2016. [Online]. Available: http://googleresearc​h.blogspot.com/2015/​07/deepdream-code-ex​ample-for-visualizin​g.html. [Accessed: 09- Feb- 2016].
    • [2] S. Russell, P. Norvig and E. Davis, Artificial intelligence. Upper Saddle River, NJ: Prentice Hall, 2010.
    • [3] M. Nielsen, "Neural Networks and Deep Learning", Determination Press, 2015.
    • [4] N. Fraser, "Neural Network Follies", Neil.fraser.name, 2016. [Online]. Available: https://neil.fraser.​name/writing/tank/. [Accessed: 08- Feb- 2016].
    • [5] A. Mordvintsev, "Inceptionism: Going Deeper into Neural Networks", Research Blog, 2016. [Online]. Available: http://googleresearc​h.blogspot.com/2015/​06/inceptionism-goin​g-deeper-into-neural​.html. [Accessed: 08- Feb- 2016].
    • [6] Cs.stanford.edu, "Neural Networks - Future", 2016. [Online]. Available: https://cs.stanford.​edu/people/eroberts/​courses/soco/project​s/neural-networks/Fu​ture/index.html. [Accessed: 09- Feb- 2016].
    • [7] R. Sharma, "Potential applications of artificial neural networks to thermodynamics: vapor–liquid equilibrium predictions", Sciencedirect.com, 2016. [Online]. Available: http://www.sciencedi​rect.com/science/art​icle/pii/S0098135498​002816. [Accessed: 09- Feb- 2016].