Integrating causal cognition, concepts and decision-making
Imagine yourself hungry and in a new (wild) environment. To find food, you need to determine which objects might be food, observe other animals' eating habits to determine what is possibly safe, and then make decisions--possibly involving hard tradeoffs--about what exactly to eat. And of course, this whole process is dynamic: you learn from the outcomes of your choices, revise your understanding of what is edible, and so on. Everyday cognition similarly requires an astonishing range of cognition, even in familiar environments: we categorize objects, make inferences about them, learn causal relations between them, make plans and take actions given causal beliefs, observe and categorize the outcomes of our actions, learn from those outcomes, and so on. Despite this variety, our cognition is quite seamless: we move between cognitive processes and representations with ease, and different types of cognition seem to share information easily. This integration is unsurprising, since an isolated cognitive process or representation will almost always be either highly specialized, or useless.
This picture of integrated cognition stands in stark contrast with much of the practice of the cognitive sciences. While allusions are often made to the relevance of other cognitive operations, research almost always focuses on a single process or representation, particularly in the cognitive domains of: (i) causal learning/reasoning; (ii) concept acquisition/application; and (iii) decision-making/action selection. Experiments on causal learning rarely require decision or action in the world; concept learning experiments focus on judgments about likelihoods that novel objects fall in various categories; decision-making experiments explicitly provide relevant probabilities; and so on. The present project aims to bring these three areas of cognitive science together by: (a) developing integrated models of this particular trio of cognitive domains; and (b) determining how those models are, or could be, implemented in neural systems or circuits.
The first component of this project aims to develop and empirically test an integrated model of causal, conceptual, and decision-making cognition. This model will provide a single explanatory framework, and thereby help unify our theoretical and empirical understanding of these three cognitive domains. This integrated model will be developed using the framework of graphical models: a rich computational framework that provides the representational and algorithmic resources to capture much of our complex cognition, and has already proven successful for several cognitive domains. The second component of the project will explore lower-level representations and processes that could plausibly implement the high-level, integrated model of these three different cognitive domains. Rather than trying to jump from high-level cognition to neural systems in one move, two different strategies will be pursued in parallel: (i) developing implementations using lower-level cognitive models for which there are plausible neural implementations; and (ii) investigating structured neural networks that can capture learning and inference in dynamically changing graphical models. The former will provide an intermediate stage in a multi-step implementation; the latter will provide theoretical guidance about how neural implementations are possible for the high-level integration of the three cognitive domains.