In the spirit of Bayesian Enlightenment, as suggested by Jones & Love (J&L), grounded cognition offers architectural mechanisms that naturally afford Bayesian analysis. In particular, the constructs of simulator and situated conceptualization illustrate the potential for integrating explanations across Marr's (Reference Marr1982) computational, algorithmic, and implementation levels. Many other grounded constructs also undoubtedly offer similar potential.

In perceptual symbol systems, a simulator is a dynamical system distributed across the modalities that process a category's properties, aggregating information about diverse instances, comparable to a concept in traditional theories (Barsalou Reference Barsalou1999; Reference Barsalou2003a). The beer simulator, for example, aggregates information about how beer looks, smells, and tastes, how it is consumed, how we feel afterwards, and so on. If someone experiences a wide variety of beers (e.g., American, Belgian, English, German, Czech, Indian, Thai, etc.), the beer simulator captures diverse multi-modal states about the category (modeled naturally with neural net architectures; e.g., Pezzulo et al. Reference Pezzulo, Barsalou, Cangelosi, Fischer, McRae and Spivey2011). On a given occasion, the beer simulator dynamically produces one of many specific beer simulations, from an infinite set possible. A natural way of thinking about the space of possible simulations within a simulator is as a space of Bayesian priors, with some simulations being more likely than others. Furthermore, the strength of a given prior can be assessed empirically (rather than simply assumed), reflecting well-established and readily measured factors, including how frequently and recently category instances have been experienced, how ideal or preferred they are, their similarity to other instances, and so forth (Barsalou Reference Barsalou1985; Reference Barsalou and Neisser1987).

In grounded approaches to the conceptual system, a situated conceptualization is a situation-relevant simulation from a simulator embedded in the representation of a likely background situation (Barsalou Reference Barsalou2003b; Reference Barsalou, Robbins and Aydede2008c; Yeh & Barsalou Reference Yeh and Barsalou2006). One situated conceptualizations of chair, for example, represents a chair on a jet, embedded in a jet setting, accompanied by relevant actions and mental states. A natural way of thinking about a situated conceptualization is as representational structure that captures and produces Bayesian likelihoods. Entering a jet setting, for example, may activate the situated conceptualization for jet chairs, producing the expectancy that this specific type of chair will be experienced shortly. Similarly, seeing a jet chair activates expectancies about how to interact with it, how it will feel to operate, and so on. Again, the statistical structure of situated conceptualizations can be assessed empirically, through objective assessments of the environment, subjective estimates of co-occurrence, et cetera. In general, much evidence demonstrates that situated conceptualizations produce part-whole inferences to support diverse forms of reasoning (e.g., Barsalou et al. Reference Barsalou, Niedenthal, Barbey, Ruppert and Ross2003; Reference Barsalou, Barbey, Simmons and Santos2005; Wilson-Mendenhall et al. Reference Wilson-Mendenhall, Barrett, Simmons and Barsalou2011).

Together, simulators and situated conceptualizations naturally produce Bayesian inferences. Before entering a building for the first time, priors associated with the chair simulator produce situation-independent inferences about chairs likely to be encountered (e.g., kitchen chairs, easy chairs, office chairs), whereas likelihoods emerging from relevant situated conceptualizations produce inferences about likely chairs to be found in this particular context (e.g., living rooms of Buddhist friends). From the perspective of grounded cognition, priors and likelihoods are combined to produce simulations that prepare agents for what is likely to exist in the world, for how to act on the world, and for the mental states likely to result (Barsalou Reference Barsalou2009; Barsalou et al. Reference Barsalou, Breazeal and Smith2007). Because such simulations utilize the modalities for perception, action, and internal states, representations in these modalities become primed, thereby facilitating expected interaction with the environment. Although architectures remain to be developed that combine these sources of Bayesian information to produce simulations, neural net architectures that reactivate previously experienced states have much potential for doing so.

Because simulators and situated conceptualizations occur in nonhumans, they offer a natural account of conceptual processing across species (Barsalou Reference Barsalou2005). If so, the kind of Bayesian analysis just described applies comparatively, perhaps via somewhat common forms of optimality arising continuously across evolution. Where humans are likely to differ is in the linguistic control of this architecture, with words activating simulators, and larger linguistic structures specifying situated conceptualizations compositionally and productively (Barsalou Reference Barsalou1999; Reference Barsalou, Semin and Smith2008b).

Bayesian analysis can also be applied to linguistic forms, similarly to how it can be applied to simulators and situated conceptualizations. On activating a word, the probability that other words become active reflects a distribution of priors over these words, constrained by likelihoods, given other words in the context. As research shows increasingly, the statistical structure of linguistic forms mirrors, to some extent, the structure of conceptual knowledge grounded in the modalities (e.g., Andrews et al. Reference Andrews, Vigliocco and Vinson2009; Barsalou et al. Reference Barsalou, Santos, Simmons, Wilson, De Vega, Glenberg and Graesser2008; Louwerse & Connell Reference Louwerse and Connell2011). Of interest is whether similar versus different factors optimize the retrieval of linguistic forms and conceptual knowledge, and what sorts of factors optimize their interaction.

Finally, the grounded perspective assumes that cognition relies inherently on the body, the physical environment, and the social environment, not just on classic cognitive mechanisms (Barsalou Reference Barsalou2008a). Because cognition does not occur independently of these other systems, characterizing their structure is essential, analogous to the importance of characterizing the physical environment in Bayesian analysis.

For all these reasons, grounded cognition offers a natural approach for practicing and achieving Bayesian Enlightenment. As cognition emerges from bodily and neural mechanisms through interactions with physical and social environments, numerous forms of optimization undoubtedly occur at many levels. Fully understanding these optimizations seems difficult – not to mention unsatisfying – unless all relevant levels of analysis are taken into account. Indeed, this is the epitome of cognitive science.