I started exploring Google Deep Dream Art back in summer 2015, and I was fascinated. As an artist (photography and digital art, mainly Fractals) with a background in IT, I really enjoy creating those surreal works.
Trippy CarDogs and eyes everywhere, must be a dream...
When I was asked to do a commission for a cover image for 'Communications of the ACM' (a computer science magazine), using the Google Deep Dream algorithm, my answer was a loud and clear yes!, especially when the Art Director told me the topic for the story: the Google TPU, a new computer chip, designed for machine learning (neural networks!).
Many people interested in IT may know Moore's Law, the observation that the number of transistors on a chip doubles every year, respectively every two years (named after Gordon Moore, the co-founder of Fairchild Semiconductor and Intel). In simple words: the processing power of computers will double every two years. Sounds great? Yes, and it worked for some decades but unfortunately, it ended some years ago. No need to worry though, clever scientists and engineers researched new technologies like domain-specific architectures. Now, what does this mean? Well, there are plenty of long articles about this but let's put it into one word: specialization. A technology (hardware, software, both) that can do specific things really good -- and really fast.
The Google Tensor Processing Unit (TPU) is a good example of a domain-specific architecture: a chip designed for machine learning applications such as neural networks. The Google Deep Dream algorithm uses such a neural network to find and enhance patterns in images so it seemed natural to treat a photo of a TPU with a dose of deep dream.
It is hard to predict what the AI will find in a specific image which means the creative process of making a great cover image is both challenging and time-consuming. And wonderful! Looking at the results often made me laugh, how about this dog-like creature sucking at a pipe of the TPU:
Dog sucking on TPU pipeSurreal Detail, Google Deep Dream Art
I used different layers and settings to find out which of the results are worth a deep dive with different parameters. The neural network works with so-called layers with strange names like inception_4c-3x3_reduce or inception_4b-pool-proj. From my experience, I had a rough estimate of which layer would produce cute dogs or strange creatures but the algorithm often has its own ideas.
With lots of dogs, eyes and pagodas it is quite easy to overdo things and produce really trippy results. Nothing against some nice surrealism but it was important that the TPU is still recognizable as a computer chip. So I knew I wanted some dogs and some trippy elements because this is what Deep Dream is known for.
Cute DogsGoogle TPU processed with a Deep Dream treatment (detail)
Once I found some results I really liked I started tweaking the settings and combining the elements. Creativity, versatility and patience were the main elements of this commissioned work.
I really loved this challenge and I'm very pleased with the outcome:
CACM Cover 09-2018Cover Illustration by Matthias Hauser (for the article 'A Domain-Specific Architecture for Deep Neural Networks'
Image Credit for the picture above: Association for Computing Machinery, CACM Cover September 2018 Issue
Special thanks to Margaret Gray, Associate Art Director at Andrij Borys Associates, LLC in New York. It was a real pleasure for me to work with her.
Yes, sure. If you are interested in an individual Deep Dream artwork please don't hesitate to contact me.
You bet! Enjoy my wall art collection of surreal and trippy images, posters and prints (canvas, metal, acrylic) in many sizes available.
January
February
March
April
(3)
May (3)
(7)
June (7)
(5)
July (5)
August
(2)
September (2)
(2)
October (2)
(3)
November (3)
(5)
December (5)
|
(1)
January (1)
(1)
February (1)
(2)
March (2)
April
(3)
May (3)
June
(1)
July (1)
August
(1)
September (1)
October
(1)
November (1)
December
|
January
February
(1)
March (1)
(1)
April (1)
(1)
May (1)
(2)
June (2)
(2)
July (2)
(2)
August (2)
(1)
September (1)
October
(1)
November (1)
(2)
December (2)
|
January
February
March
April
May
June
(1)
July (1)
(1)
August (1)
September
October
(2)
November (2)
(1)
December (1)
|
(2)
January (2)
(3)
February (3)
(2)
March (2)
(1)
April (1)
(2)
May (2)
(2)
June (2)
July
(1)
August (1)
(1)
September (1)
(1)
October (1)
November
(1)
December (1)
|
January
February
March
April
May
June
July
(1)
August (1)
September
(1)
October (1)
November
December
|
January
February
March
April
May
June
July
August
September
October
November
December
|
(1)
January (1)
(1)
February (1)
(1)
March (1)
(2)
April (2)
(1)
May (1)
(2)
June (2)
July
August
September
October
November
December
|
January
February
March
April
May
June
July
August
September
October
November
December
|
January
February
March
April
May
June
July
August
September
October
November
December
|