CCN Colloquium: "Facing Down the Interface Problem: No Distinctive Cost to Representational Diversity"

Abstract: Mental states represent the world in systematically different ways, encoding contents in multiple "codes" or "formats." For example, some mental states are thought to be image-like, others sentence-like; some analog, others symbolic. The interface problem concerns how agents can coordinate contents across different mental codes. The standard assumption is that interfaces incur distinctive costs: translation from one format to another is explanatorily mysterious and computationally difficult. I argue that there is no reason to think that between-format representational transitions, or "translations," are distinctively more challenging to explain or to execute than within-format transitions, or "inferences." In fact, translation and inference can exemplify similar properties and fulfill similar functions. Where classical computational models used inference to solve cognitive tasks, many contemporary machine learning models use translation. The diversity of representational formats is not so much a barrier to rational agency as it is a way of realizing it.