Training of artificial neural networks allowed us to make a Christmas story
Over the past 10 years, the training of artificial neural networks has made significant progress. AI is already allowed to create independently, although it turns out not always good. For example, in 2018, Janel Shane and the New York Times tried to develop Halloween costumes using a bot. The works written by the machine can only be used as entertainment, but it is early to release a collection of stories. This is confirmed by the experiment conducted by Professor Will Knight. A scientist at the Massachusetts University of Technology entrusted the robot with writing a Christmas fairy tale. It turned out so and so, mildly speaking.
How the training of artificial neural networks for the story was held
The researcher applied textgenrnn algorithm, used to create Halloween costumes. The team prepared a sample of 360 themed films and copied the descriptions from Wikipedia. Will Knight said that this version of AI is able to either generate letters into words, or collect from words separate sentences.
The program provides dozens of parameters, but a group of professionals interested only three.
- “Number of layers”. The indicator characterizes the complexity of the neural network. The more of them are, the higher the complexity of the information that artificial intelligence can process.
- “The number of epochs.” The machine accesses the source data strictly a certain number of times, this is customizable by a person. Such appeals are called epochs.
- “Temperature”. The lower the selected value, the more often the robot will turn to common words and phrases. When raised, the neural network will apply expressions that are used less frequently.
Unfortunately, the results were not impressive. The text summoned by the computer turned out to be quite strange. A lot of inconsistencies and problems with punctuation was revealed there. Punctuation marks were missing or inserted incorrectly. Artificial intelligence generated such “pearls”, “tennis Christmas discovery”, “dead California”, “The janitor injured Christmas”, etc. Alas, there was a complete nonsense.
Professor Knight commented on the negative outcome of the experiment. He saw the first reason in the insufficient amount of information used in the course of training. According to the researcher, 360 descriptions are not enough for correct work. Also textgenrnn was not that good in building phrases compared to other programs. The biggest difficulty lies in the incorrect functioning of the mechanism of generating offers in the service. More high-quality information could be taken, but it unlikely would help to obtain an agreed text.
However, individual synopses turned out to be very funny. After studying, the project team allocated them and handed them to the specialists who made posters. For example, there were such curious names as “Fighting Christmas” or “Christmas Mile”. Surely many would go to the cinema for the premiere of such a film.
Artificial intelligence has already been trusted to write fairy tales. For example, the community of online writers Botnik published a story prepared by AI based on the works of the brothers Grimm. Unfortunately the work “Princess and Fox” also turned out strange. The main problems are the same: inconsistency of sentences, absence and incorrect punctuation marks inserting. Earlier users presented a project composed by a robot, focusing on the books of Joan Rowling about Harry Potter.