School of Communication and Culture

This article investigates the unpredictable vector of liveness [1] within the context of machine learning art with a focus on text-based processing. It is observed that there are similarities between generative art and machine learning art as both produce unpredictable results. According to Noah Wardrip-Fruin [2], the generative art form, such as Loveletters (1952), can be considered as a system that generates unpredictable outcomes. Loveletters, allegedly the first digital literary work by computer scientist Christopher Strachey, is regarded as an ‘unpredictable manifestation’ of a system [3]. This system generates different variations of love letters, and such unpredictable manifestation is conditioned by two hidden elements: data and processes [3]. The use of random algorithms plays an important role in generative art (Turing’s random algorithm with its’ random number generator was used in Loveletters) to produce autonomous and unpredictable outcomes. However, machine learning emphasizes ‘predictive power’ [4], in which prediction is produced through feeding in a large amount of training data. Additionally, this kind of system employs predictive models and statistical algorithms to accomplish data processing and analysis. Machine Learning Art, such as text/novel generators, is claimed to be able to produce text with the similar writing style of the provided training corpus, but it also produces unpredictable text through setting different control parameters, such as number of epochs, amount of neural network layers and their hidden units, temperature and batch size.

This article is the result of the experiment of an open source machine learning library called ml5.js [5], which is built on top of TensorFlow.js, a Javascript framework, for training and deploying machine learning models. ml5.js provides immediate access in the web browser to pretrained models for generating text. A Python training script employs the tensorflow library, which is used in the ml5.js environment to take in a large amount of text, and train a custom dataset as a pretrained model [6]. The study of the javascript libraries and the python script, with a specific focus on next character prediction and recurrent neural networks (RNN), unfolds the machine learning processes from data training to Long Short-Term Memory networks [7][8]. Building upon the notion of generativity [9][10][2], this article discusses the (un)predictable vector by examining the intertwining force between predictability and unpredictability that constitutes the liveness of text-based processing in machine learning art. This paper argues that the (un)predictable vector of liveness helps to build an understanding of the relation between, but not in separation, training and execution processes, as well as the resultant actions that extends the aesthetic and live experience of machine learning art. The article contributes to the border understanding of generativity and liveness in machine learning art that employs generative models.