Interrogating the Ontological Assumptions in Deep Learning Architectures

Abstract

One of the lessons that the Turing Test taught the community of intelligence practitioners was that the systemic architecture of a cognitive system dictates, among other things, the form of the output, the kind of knowledge that can be had about the system, the kinds of knowledge manipulation that are possible within the system, and the epistemological and ontological commitments that are implicit in the design of those architectural elements. This idea has received increased attention of late because we have come to recognize that the ethical implications of systems stem from their decision-space and that systems have more than just operational consequences. That is, we have grown to understand that not only does architecture matter: it matters a whole lot! The degree to which theories of mind supported by archetypes such as deep learning architectures accurately reflect the nature of human (natural and artificial) intelligence has also been a well-investigated research topic of cognitive science. [1]One of the most promising conclusions of cognitive science, as it relates to the design of artificial intelligence systems, is that cognitive systems are profoundly ecological agents. That is to say, the form of a cognitive architecture is deeply and essentially tied to the form of a physical architecture whereby that cognitive agent is situated in an ecosystem that provides the physical resources required by the agent to enact intelligent agency. As we seek design principles, working architectural models, and ethics-in-architecture for artificial general intelligence, it is therefore vital to explore the ontological commitments associated with different forms of cognition-engaging architectural elements. Such interrogations are also essential if we are to mitigate the influence of implicit bias in the data that we use to refine our deep learning systems.