The social lives of generative adversarial networks

Abstract

Generative adversarial networks (GANs) are a genre of deep learning model of significant practical and theoretical interest for their facility in producing photorealistic 'fake' images which are plausibly similar, but not identical, to a corpus of training data. But from perspective of a sociologist, distinctive architecture of GANs is highly suggestive. First, a convolutional neural network for classification, on its own, is (at present) popularly considered to be an 'AI'; and a generative neural network is a kind of inversion of such a classification network (i.e. a layered transformation from a vector of numbers to an image, as opposed to a transformation from an image to a vector of numbers). If, then, in training of GANs, these two 'AIs' interact with each other in a dyadic fashion, shouldn't we consider that form of learning... social? This observation can lead to some surprising associations as we compare and contrast GANs with theories of sociologist Pierre Bourdieu, whose concept of so-called habitus is one which is simultaneously cognitive and social: a productive perception in which classification practices and practical action cannot be fully disentangled. Bourdieu had long been concerned with reproduction of social stratification: his early works studied formal public schooling in France not as an egalitarian system but instead as one which unintentionally maintained persistence of class distinctions. It was, he argued, through cultural inculcation of an embodied and partially unconscious habitus---a durably installed generative principle of regulated improvisations---that, he argued, students from upper classes are given an advantage which is only further reinforced throughout their educational trajectories. For Bourdieu, institutions of schooling instill deeply interiorized master patterns of behavior and thought (and classification) which in turn direct acquisition of subsequent patterns, whose character is determined not simply by this cognitive layering but by their actual use in lived practice, especially early in childhood development. In this work I develop a productive analogy between GAN architecture and Bourdieu's habitus, in three ways. First, I call attention to fact that connectionist approaches and Bourdieu's theories were both conceived as revolts against rule-bound paradigms. In 1980s, Rumelhart and McClelland used a multilayer neural network to learn phonology of English past-tense verbs because sometimes we don't follow rules... language is full of exceptions to rules; and in case of Bourdieu, habitus was an answer to a long-standing question: how can behaviour be regulated without being product of obedience to rules? Bourdieu strove to transgress what was then seen in social sciences as a conceptual opposition between structure-based theories of social life and those which emphasized an embodied agency. Second, I suggest that concerns about bias and discrimination in machine learning in recent years can in part be attributed due to increased use of ML models not just for static classification but for practical action. Similarly, habitus for Bourdieu is simultaneously durable and transposable: its judgments may be relatively stable, but are capable of being deployed dynamically in novel and varying social situations---or what ML practitioners might call generalizability. We can thus theorize generative models (including GANs) as biased not just in their stereotyped classifications, but through their potential for actively generating new biased data. These generated actions then recursively become part of social arena Bourdieu called field, into which new agents are 'born' and for which they may know few alternatives. Finally, it is intriguing that GAN researchers and Bourdieu both extensively use metaphors from game theory. Goodfellow described GAN architecture as a two-player minimax game with function V(G,D), meaning that there is a single abstract function whose output discriminator is trying to maximize and which generator is trying to minimize; but dynamic nature of GAN training process means that convergence to Nash equilibrium is nontrivial. But for Bourdieu, such a utility-based approach to artistic creation could not be more crude when compared to social reality of art worlds: utilitarianism is, for him, the degree zero of sociology, by which he means an isolated, inert, and amodal---and therefore not particularly sociological---starting point. Moreover, 19th-century bohemian culture was characterized primarily by its inversion of financial incentives, in which failure is a kind of success, and selling out (i.e. maximizing profit) worst of all; and thus relentless optimization of neural networks may be fundamentally at odds with value functions of many human artists. I conclude that deep learning, while primarily understood as a scientific and technical achievement, may also intentionally or unintentionally constitute a nascent, independent reinvention of social theory.