deep neural networks

Click on the image to open it in a modal lightbox

This diagonal array of Necker cubes has been processed in an iterative fashion with the Google DeepDream algorithm which has been implemented in Python. A code snippet can be found below and the full source code is available on GitHub.

https://github.com/google/deepdream

An interesting question is: If an AI creates a work of art, who owns the rights to it?

Hart, R.. (2017). If an AI creates a work of art, who owns the rights to it?

Show/hide publication abstract

“IP + ai if an ai creates a work of art, who owns the rights to it? a robot arm in trafalgar square, london, replicates a drawing being created by austrian artist alex kiessling from vienna, with another robot arm simultaneously replicating the work in berlin, thursday, sept. 26, 2013. the project has been given the title ‘long distance art’ and utilizes the irb 4600 robotic arms, produced by the company abb for industrial manufacturing in areas such as automotive, plastics, metal fabrication and electronics. (ap photo/matt dunham) beauty is in the active pixel sensor of the beholder. (ap photo/matt dunham) share written by robert hart obsession machines with brains august 15, 2017 artificial intelligence is already capable of creating a staggering array of content. it can paint, write music, and put together a musical. it can write movies, angsty poems, and truly awful stand-up comedy. but does it have ownership over what it produces? for example, an ai at google has managed to create sounds that humans have not heard before, merging characteristics of two different instruments and opening up a whole new toolbox for musicians to play around with. the company’s deepdream is also capable of generating psychedelic pieces of art with high price tags; last year two sold for $8,000—with the money going to the artists who claimed ownership over the images. screen shot 2017-08-15 at 10.55.18 am an example of the work google’s deepdream algorithms can create. (google deep dream) as it stands, ais in the us cannot be awarded copyright for something they have created. the current policy of the us copyright office is to reject claims made for works not authored by humans, but the policy is poorly codified. according to annemarie bridy, a professor of law at the university of idaho and an affiliate scholar at stanford university’s center for internet and society, there’s no actual requirement for human authorship in the us copyright act. nevertheless, the ‘courts have always assumed that authorship is a human phenomenon,’ she says. eran kahana, an intellectual-property lawyer at maslon llp and a fellow at stanford law school, doesn’t believe we should award authorship to ais. he explains that the reason ip laws exist is to “prevent others from using it and enabling the owner to generate a benefit. an ai doesn’t have any of those needs. ai is a tool to generate those kinds of content.” he likens the idea to a computer word processor using spell check. if you make a s…”

def objective_L2(dst):
    dst.diff[:] = dst.data 

def make_step(net, step_size=1.5, end='inception_4c/output', 
              jitter=32, clip=True, objective=objective_L2):
    '''Basic gradient ascent step.'''

    src = net.blobs['data'] # input image is stored in Net's 'data' blob
    dst = net.blobs[end]

    ox, oy = np.random.randint(-jitter, jitter+1, 2)
    src.data[0] = np.roll(np.roll(src.data[0], ox, -1), oy, -2) # apply jitter shift
            
    net.forward(end=end)
    objective(dst)  # specify the optimization objective
    net.backward(start=end)
    g = src.diff[0]
    # apply normalized ascent step to the input image
    src.data[:] += step_size/np.abs(g).mean() * g

    src.data[0] = np.roll(np.roll(src.data[0], -ox, -1), -oy, -2) # unshift image
            
    if clip:
        bias = net.transformer.mean['data']
        src.data[:] = np.clip(src.data, -bias, 255-bias)

def objective_L2(dst):

dst.diff[:] = dst.data

def make_step(net, step_size=1.5, end='inception_4c/output',

jitter=32, clip=True, objective=objective_L2):

'''Basic gradient ascent step.'''

src = net.blobs['data'] # input image is stored in Net's 'data' blob

dst = net.blobs[end]

ox, oy = np.random.randint(-jitter, jitter+1, 2)

src.data[0] = np.roll(np.roll(src.data[0], ox, -1), oy, -2) # apply jitter shift

net.forward(end=end)

objective(dst) # specify the optimization objective

net.backward(start=end)

g = src.diff[0]

# apply normalized ascent step to the input image

src.data[:] += step_size/np.abs(g).mean() * g

src.data[0] = np.roll(np.roll(src.data[0], -ox, -1), -oy, -2) # unshift image

if clip:

bias = net.transformer.mean['data']

src.data[:] = np.clip(src.data, -bias, 255-bias)

References

Mordvintsev, A., Olah, C., & Tyka, M.. (2015). Research Blog: DeepDream – a code example for visualizing Neural Networks

Plain numerical DOI: 10.1111/j.1600-6143.2012.04033.x.The
DOI URL
directSciHub download

Show/hide publication abstract

Dean, J., & Monga, R.. (2015). TensorFlow – Google’s latest machine learning system, open sourced for everyone

Show/hide publication abstract

“Deep learning has had a huge impact on computer science, making it possible to explore new frontiers of research and to develop amazingly useful products that millions of people use every day. our internal deep learning infrastructure distbelief, developed in 2011, has allowed googlers to build ever larger neural networks and scale training to thousands of cores in our datacenters. we’ve used it to demonstrate that concepts like ‘cat’ can be learned from unlabeled youtube images, to improve speech recognition in the google app by 25%, and to build image search in google photos. distbelief also trained the inception model that won imagenet’s large scale visual recognition challenge in 2014, and drove our experiments in automated image captioning as well as deepdream. while”