The deep recurrent neural network used in this work tries to learn a complex mathematical model of the probabilistic distribution of sequences of characters. The term “recurrent” refers to the fact that some of the neurons inside the network feed back into the network’s inputs, allowing it to keep the past characters read in its memory. As it is trained, it gets a sense of how certain characters follow one another, from groups of 2-3 letters that correspond to morphemes and on to syllables, words, and eventually whole sentences. Because it is trained at the character level, the system can generate non-English words. None of the words, syntactic structures, or sentences found in the book were in any way explicitly encoded in the system by the artist who developed it: Emily Brontë’s novel is, effectively, the only thing at all that this deep recurrent neural network knows about language, or indeed the world.
The work is released as a limited series of 31 unique artbooks. Each book is signed by the author.
The book was designed by the editor. When “chapter” followed by some letters occurs as its own line, this line was typeset as a chapter heading on a new page. No changes were made to the output.
Audry, Sofian. “Unrolling the Learning Curve: Aesthetics of Adaptive Behaviors with Deep Recurrent Nets for Text Generation”, International Symposium on Computational Media Art (ISCMA) 2019, Hong Kong, 2019.
Brontë, Emily. Wuthering Heights, Project Gutenberg, 1996/1847.