In the world there are already a lot of algorithms that are based on neural networks can generate a variety of works. They are still a little applicable in practice, and some are used for fun. Last year, Janelle Shane, together with the New York Times used AI to generate the names of Halloween costumes. And recently, will knight at MIT decided to repeat the experiment, only in this case, the same AI is used in order to create stories Christmas stories. It turned out very ambiguous.
For work, it was decided to select 360 Christmas movies, descriptions of scenes which has been taken from Wikipedia. The algorithm used was the same as when generating the names of Halloween costumes textgenrnn.
“Textgenrnn has two modes: you can use the letters to generate words in the style of other words, or use words to generate sentences in the style of other proposals.” — explain the principle of operation of AI will knight
The neural network has a lot of options, but the researcher concentrated on three of them — “number of layers”, “number of epochs” and “temperature”. This is a provisional name and they reflect not what you think by looking at them.
“Layers are related to the complexity of neural networks: the more layers it has, the harder data it can handle. Number of epochs is the number of times that there is at neroli to see the original training data, after which it can produce results. Well temperature: the lower it is, the more the network will choose common words in the dataset compared with those used less often.”
In the end, all the results looked very strange. Here is one synopsis. The translation is very approximate, because the neural network wrote sometimes inconsistent with each other words and did not comply with the spelling and punctuation.
“Mary needs to find friends from magic and to stop Nathan from the beginning of the Christmas. Serious resort elves have taken care of that with a tennis Christmas opening to dead California, where the Groundhog is waiting, as the wise janitor injured Christmas. The heroes need to survive in this show no matter what.”
Sounds like nonsense and is not even fun, unlike last year’s experiment, when AI offered a very funny names Halloween costumes like “sexy man of Minecraft”. So I thought, and the researcher started to understand the problem.
“One of the problems was that the neural networks are given too little training data. Only 360 pieces. Usually to train the neural network should be several times more information. Another problem was that textgenrnn was not so good in constructing sentences compared to the alternatives. But the main reason is the limitations of the generation proposals. Even if I were to use more data and they would be better quality, to achieve coherence of the description is practically impossible.”
Although some sane and even funny results to achieve the same success. Some of the synopses turned out quite funny.
- Family Christmas terrorist moves to a new town and plans to celebrate Christmas.
- Princess-Ogre wants to cancel Christmas
- Angel Santa Claus works in a factory
Some names of non-existent movies turned out pretty good and they even drew the posters.