Understanding how AI functions, to create and succeed

How the artificial intelligence thinks:

At the turn of the 21st century, Windows XP was released, Windows 95 would become a legacy OS and IBM released the first 64-bit operating system. Only three years beforehand, in 1997, Deep Blue, an IBM designed computer beat the world champion at Chess, this was a pivotal moment for computing history - the dawn of artificial intelligence mastering human skills. The excitement around the success of AI during this period led to films like 2001's Space Odyssey.

AI development has not slowed down, from iRobot's Roomba - intelligently mapping homes across the nation, to the first autonomous car by Google. During and beyond the 2000's we have and will continue to see rapid developments in AI technology. To briefly understand how today's AI works to develop art and within other sectors we can look at DeepMind's AlphaStar.

It's an AI designed to beat the top professional player in a game called Starcraft II. Starcraft II is highly regarded as being one of the most micro (small) and macro (large) gameplay intensive games requiring a large amount of actions per minute to be good at the game. Up to 500 at some points in the game for the best players.

This AI hasn't yet completed its end game goal of defeating the best player in the game, but it has beat several top and very competitive ones along its journey. Now disguised under anonymity, players can face it online at any time during their match making. It uses a detailed process to better itself and has become at least better than 99.9% of players.

To understand this process, I learnt about AI evolution, AlphaStar was fed a number of initial games/replays to understand how humans play the game. The AI then learnt this and played against itself. Each iteration led to either a loss or a win, this was the end game conditioner or success rate.

Among other things, the AI would also measure its ability to use the economy of the game to better itself, its kill to death ratio among other things. As you can see with the above video excerpt, this results in an outcome prediction - the higher that is, the better of a chance it has at winning.

However, one problem is that the AI up to a point only played against itself and by seeing other humans play, amongst other variables, it couldn't adapt to ongoing trends in the multi-player scene. So DeepMind, the creator of AlphaStar started having it play online versus other players.

By playing lots of players at once, and employing different strains of the same machine, it could multi-variably compare its win rate among other metrics and select the best version of itself to go against the next opponent.[9]

Through this process the AI has demonstrated various different methods at problem solving, both divergently and convergently - typical characteristics of the creative individual.

As you can see, AI training techniques have developed to such a stage where biases can be reduced significantly and moreover, human replication and mastery in skills can be achieved. Using this very same process, an AI can learn how to paint, create complex scenery, it can be used to edit movies, among other novel use cases.[30]

Creative merit test:

Creative merit test: In the interests of understanding whether an AI can be creative, we can look at a series of pictures. These pictures will have a human drawn piece of art and one by an AI, your goal is to find out whether or not you can tell the difference and to also rate which one you feel has the most creative merit within the limits of the prescribed definition. Pick one piece of art, either right or left which you think was created by an AI.

Here's the definition of creativity for your recollection:

"The quality to be able to create new and unique ideas or look at a problem through a new perspective. This process ends up creating some form of value."

The left picture was created by an AI, the right was by a human.

The left picture was created by an human, the right was by a AI.

Finally, the left picture was created by an human, the right was by a AI.

Ultimately it's up to the individual to decide, I would say these images are very creative, they form new and unique shapes with an equivalent contextual value to something you'd see in any art gallery. In fact some machines have already made their way to art galleries.


The Portrait of Edmond Belamy:

This artwork named "Portrait of Edmond Belamy" sold for a total of $432,500 at Christie's auction house. Not only is this a fine example of how a machine can add value to the artwork its creating but offers great insight into how a machine thinks. This particular piece was created using a generative adversarial network.

A generative adversarial network is best described using this diagram:

The random noise is just a variable inputted into the generator, which is generated at first and for examples sake, is some static which is its attempt at an initial realistic generated image. It then goes from the generator into a discriminator which looks at whether or not the image is realistic.

The generators goal is to give out a huge amount of attempts at creating realistic artwork and the discriminators goal is to reduce those attempts to those which are fake and those which are real based on a training set.

Within the configuration of the generator, it looks at what was rejected by the discriminator and it's probability for rejection, so for example a generated portrait with a blue hat won't look like those tagged as realistic within the source material (within these the realistic tagged images they all wear black hats).

Therefore it knows that blue in this configuration above the head of the subject makes for a reduction in the probability for success. But maybe the colour of skin of the person in the portrait is more accurate, which would increase it's probability. Eventually through multiple iterations and extremely fast A/B testing. The AI can find the most realistic outcome.[33]

This kind of divergent and convergent decision process is very easy to make comparisons with common creative thought processes. What more, that same process can create pieces of art of extraordinary value.

Now imagine this kind of machine but at a larger scale, we have Google with it's vast array of classified images and robot scanners which act as active viewers on websites that collect information from them as needed for analytics and search indexing.[32]

Combine those two together and you have the source material for a generative adversarial network the likes you've never seen before that can create things most people couldn't tell were real... This makes for a great introduction to what I will explore in the final chapter.

In conclusion however, we've discovered that today the extent of the development of artificial intelligence is such that it can be creative in terms of the widely accepted definition, with levels of independence achieved through neural networks that can be predicted but are still innately unique and based on randomised data.

Maybe you don't believe based on your experience today that these paintings or products hold real creative value, but to some the value of these paintings is timeless and highly collectable.

The future it seems doesn't look very human after all.