Publication Alert: Out now in Entropy, a paper from IMC researchers proposes a new way of thinking about how minds carve up an ambiguous world into meaningful categories.
"It Is What It Isn't: Introducing a Constraint-Based Approach to Structure Learning" by Christoffer Lundbak Olesen, Nace Mikuš, Mads Hansen, Nicolas Legrand, Peter Thestrup Waade, and Christoph Mathys argues that structured representations don't need to be built through explicit inference over predefined possibilities. Instead, they can emerge from an ensemble of simple learning processes that continuously compete and cooperate under system-level constraints.
Using a population of individual inference processes, each a Hierarchical Gaussian Filter (HGF), the model shows how stable category structure can self-organise through constraint-based dynamics. HGFs are added, strengthened, weakened, and removed over time in response to incoming observations, allowing the model to adaptively expand, contract and reorganise its system-level structure.
The paper uses a hyper-simplistic simulation environment and introduces the model as a proof-of-concept for a novel approach to computational modelling of cognition. As such, the contribution is theoretical and serves as motivation for broadening the conceptual horizons of computational tools in cognitive science.
The paper is accompanied by an interactive simulation tool where you can watch the constraint-based dynamics unfold, explore different parameter settings, and see how components form and dissolve as the system learns: