"Earth Measurer": Visualizing a Recurrent Vanilla Neural Network, Figuring Abundance

• In Event: Regenerating Algorithms

Abstract

In the algorithm I strive to visualize, computational states are stored and reused, an architecture called "recurrence"; states that provide lower levels of error, or "loss," are given greater weight, in a move called "backpropagation," the most common form of which has been termed "vanilla." Thus, the algorithm in question is termed a "recurrent vanilla neural network." The overall design of neural networks is based on an abstract model of the neuronal physiology of animals; such algorithms employ "hidden" layers of nodes intervening between input and output. The discursive force of the neural network—its marketability, its political pliancy—stems from two interrelated qualities: this enigmatic "hiddenness," and a vague metaphorical relationship to natural phenomena.

“Nature,” William James said, “is but another name we give to excess.” One of modernity's normative moves figured nature as abundance; only recently have we begun to comprehend life on earth as fragile and ephemeral. Our computational systems, meanwhile, are increasingly granted the autonomy, authority, and alterity of natural phenomena.

For input data, I'm using genus names of the Geometridae, which translates as "earth measurer," a family in the order Lepidoptera, the butterflies and moths. The algorithm's output, strings of characters generated by iterative statistical inference, take up the flavor of scientific nomenclature: Latinate morphemes, the husks of European surnames. Putting these output "names" into the form of taxonomic lists and zoological catalogues, I visualize the rhetorical overdetermination of machine-learning in post-natural discourse, and highlight the varieties of "loss" in which our technologies are implicated.

Author

• Matthew Battles Battles, Harvard's Berkman-Klein Center for Internet and Society