An Integrated Model of Semantics and Control, Part 2: Solving the Similarity Paradox Through Context Inference

Semantic similarity plays an ambiguous role in models of human cognition. On the one hand, it is often viewed as a foundational construct that shapes how we categorize, learn, and make inferences about objects and their properties. On the other hand, a host of behavioral evidence suggests that similarity is too rigid to explain the flexibility of inductive inference. We present the Integrated Semantics and Control — Context Inference (ISC-CI) model to resolve this tension, proposing that flexible inference emerges within a system that dynamically reshapes represented semantic similarities amongst stimuli depending upon the immediate context. The ISC-CI model builds on prior models of semantics and control that learn how to build and flexibly access semantic knowledge from observing the statistical relationships between objects, their properties, and the contexts in which these occur. Critically, it introduces a new mechanism that infers a suitable representation of context for both familiar and novel scenarios, without any direct labeling in the environment. The inferred context allows the system to selectively weight different dimensions within its representational space depending on the items being processed. Through simulations and experiments, we demonstrate that the ISC-CI model provides a coherent account of performance across inductive inference and semantic similarity tasks, including classic tasks that have long challenges theories of induction, offering a unified account of these cognitive processes that highlights the importance of context. We conclude by considering the implications of these findings for broader questions in cognitive science and artificial intelligence.