Subsymbolic: The Problem of the Icon

Visiting Speaker: Sarah Pourciau (Duke University)
March 11th
5:30 PM
Gilman 479

In 1986, the members of the “Parallel Distributed Processing” research group published a volume dedicated to their new, network-based perspective on artificial and natural intelligence. They called their approach “subsymbolic.” The methods described in this publication laid nearly all the foundations for contemporary machine learning practices with one fell swoop; from a technical vantage point, therefore, they are widely known and well understood. These techniques emerged, however, from a set of hypotheses about the nature of symbolic representation that has never, to my knowledge, been systematically analyzed. I will use my talk to unpack the implications of some of these hypotheses, which turn almost exclusively around problems of meaning and interpretation: problems of the icon, in the Peircean sense, as well as of the symbol, in the strong Kantian one. I will argue that the subsymbolic is first and foremost an attempt to think through the operations of what Kant called the “productive imagination,” that the mathematics of vectorial transformation have a parasitic relationship to the categories of the aesthetic, and that a semiotic interpretation of network structure carries important de-fetishizing consequences for our understanding of  contemporary models, with their blackboxed depths, their grounding problems, and their quasi-mystical “latent spaces.”