Generative models able to mimic human creativity, or even initiate their own creative process and output in unexpected ways, without human assistance.
Creativity, the act of producing something novel and unexpected, is understood intuitively, yet challenging to formalize. Software or robots, apart from learning to be creative themselves, could help us understand and enhance this notion. Recent advancements in the field of deep learning have led to an explosion of systems that seem to exhibit a form of creativity. Some have composed a convincing pop song, learned surprising new strategies for board games without supervision or human input, and others painted images in the style of famous painters and written simple short stories and news articles.
Simple algorithms based on fixed formulas or rules such as fractals and cellular automata could lead to the emergence of complex, unexpected, and artistically pleasing results without any form of learning. The options for possible creations using one of these formulae are complicated but highly restricted, and, at best, represent a minimal form of creativity. To get closer to human creativity, an approach called generative modeling involves collecting a large amount of data in some domain (e.g., millions of images, sentences, or sounds, etc.) and then training a model to generate similar data. The achievable degree of novelty or originality for an algorithm depends on the size and variety of the dataset and the number of patterns it recognizes.
A new, dominant type of generative model, so-called generative adversarial networks, could surpass the abilities of previous iterations. For instance, it would generate a diverse number of actions like image generation, image synthesis from captions, image editing, visual domain adaptation, text generation, and many others. In this model type, two neural networks would be in constant competition with each other, one generating data and the second one judging how similar this data is to the real modeled data. It seems like the concept of competition between different agents could be a driving force for creativity. A similar mechanism is employed by modern state-of-the-art systems that beat humans at complex games such as chess, poker, and go. By letting the agents compete with themselves through self-play, they could learn creative and novel ways of beating even the most advanced human players.
Could computational systems soon be credited with authorship of artwork? Part of the difficulty in accepting a machine as being genuinely creative stems from the fact that concepts of creation, inspiration, and art are often referred to in somewhat mystical terms, something unique and primal beyond the realm of science, technology, or even rational understanding involving some form of soulfulness. Once a computational model generates something humans would judge as inspirational or creative, that piece of art could then be judged, and by knowing every little computation involved in its creation, it could paradoxically lose its very essence and inspiration.
In this scenario, questions as the following would arise: are machines able to mimic creativity, or is human creativity able to be reduced to a series of computational steps? In the face of such dilemma, maybe it would be better off by judging creativity based on its resulting works and not the mechanism by which it is produced.
If humans ever develop artificial intelligence with human-level intelligence and consciousness, it would be able to create art, since it would have the same capacity for knowledge, emotions, and social relationships. By becoming creative, computers could benefit art and artists alike, by creating new tools, modes, and styles of expression, and empowering everyday users, instead of simply substituing them.