When Images Become Logic

• In Event: Platform Practices and Predictive Seeing 2

Abstract

Machine learning presents a substantial change in both the production and operation of computational codes. Unlike the application of pre-determined logic to data that takes place in hand-coded programming, in machine learning masses of data inscribe logic through the training of neural networks. A neural networks reasoning is proving difficult to comprehend as it traverses the behaviors of interconnected layers of 'neurons'.

Matteo Pasqinelli describes the training of a neural network as a process whereby a representation of the world in a dataset becomes an operative function. In most circumstances, the generated function is tested or employed upon on similar descriptions of the world. For example, a database of digital images of numbers is used to train a network which is then used to classify other digital images of numbers. A network's 'accurate' prediction or classification is thus reliant upon the consistency of description and the world described.

The growing general application of machine learning, in less restricted settings, calls for studies into the practices and material conditions that are drawn together in the formation of its algorithms and their ongoing conduct in the world. This study takes up artistic labour with machine learning as a valuable site to do this. Examples studied display how training with specific data sets creates differences in how neural networks as operational formations in the world 'see' or 'know'. And how in a world where formations such as predictive seeing are active, images 'seen' by networks have differential productive potentialities.

Author

• Monica Monin, University of Technology Sydney