In the target article, Mitchell et al. characterize associative learning phenomena according to the relationship established between two previously unrelated stimuli (i.e., a conditioned stimulus [CS] and an unconditioned stimulus [US]). Associative learning, they suggest, occurs when the CS becomes propositionally related to the US using effortful, controlled, and rational processing. We believe this view does not account for important questions about how the representations in question (the CS and US) are modified by experience. Furthermore, this view makes assumptions about how stored representations are activated. We suggest that stimulus representations are realized by multimodal states reflecting both exteroceptive and interoceptive information brought online by a combination of top-down (e.g., propositional) and bottom-up (e.g., stimulus-driven) processes. In this view, learning occurs when the multimodal representation of a stimulus acquires an affective component. Propositional change is not necessary for learning.
A growing body of evidence suggests that the human brain captures statistical regularities in sensory-motor patterns and stores them as representations. These representations are used to continuously organize incoming sensations during the process of predicting what those sensations stand for the in world (Bar Reference Bar2003; Reference Bar2007; Kveraga et al. Reference Kveraga, Ghuman and Bar2007). External sensations always occur in a context of internal sensations from the body. As a result, the sensory-motor pattern that is stored for future use will always include a representation of the interoceptive state of the body. The brain predicts what sensations refer to in the world in part based on prior experiences of how those external sensations have influenced, or changed, internal sensations from the body on prior encounters (cf. Barrett & Bar, Reference Barrett and Barin press). These bodily changes are referred to as “affective.” Affective states can be described hedonic (pleasure or displeasure) with some degree of arousal (for a recent review, see Barrett & Bliss-Moreau, Reference Barrett and Bliss-Moreauin press). These ideas are consistent with a growing body of research demonstrating that knowledge about the world is “embodied,” or grounded, by a network of broadly distributed, diverse, multimodal states which are encoded during the experience of a given stimulus (see Barsalou Reference Barsalou2008). What you know about an object is therefore based, in part, on its affective impact in the past.
When it comes to learning, changes in a CS's meaning can be thought of as the process by which the multimodal representation of the stimulus is changed by any experience. The most fundamental change occurs because the representation of the CS is experienced in a context of affective arousal that is derived from the representation of the US. Any number of relationships between the CS and US could serve to alter the representation of the CS. The CS and US could be paired in time or space, associated semantically, or even explicitly coupled via rule-based learning. In our view, the need to differentiate types of learning in terms of how the relationship between CS and US is established (as exemplified by Mitchell et al.'s model) is eliminated. All learning can be subsumed under the same general, basic mechanism that exists in all organisms that possess the capacity to generate affective responses to stimuli in the environment. Thus, to some extent, any change in the representation of a CS is affective learning (Bliss-Moreau et al. 2008).
In typical classical conditioning paradigms, examples of USs include shocks (e.g., Vervliet et al. Reference Vervliet, Vansteenwegen, Baeyens, Hermans and Helen2005), very loud noises (e.g., Neumann & Waters Reference Neumann and Waters2006), and even sexual arousal (e.g., Hoffmann et al. Reference Hoffmann, Janssen and Turner2004). These USs act on the nervous system directly to generate a robust affective response in a bottom-up or stimulus-driven way that is automatic and unconscious. Other USs, such as negative words or pictures, have a less robust bottom-up effect on the learner's nervous system. Instead, such USs have top-down effects because they have propositional meaning. The difference in the bottom-up potency of different USs leads some theorists to believe that different models are required to account for learning phenomena. According to the affective learning perspective, this is not so – changes in affect can and do occur via both bottom-up and top-down processing and therefore with both types of USs. For example, evidence from instructed learning paradigms demonstrates that the representation of a CS can be changed by telling a person that a US will be presented after the CS, even if the US is never presented (e.g., Olsson & Phelps Reference Olsson and Phelps2004). According to the affective learning perspective, the set of instructions that indicates when the (promised, but never presented) shock will occur sufficiently alters the learner's affective state so that the interocpetive representation of this affective change is integrated into the representation of the CS. We have demonstrated that people can learn the affective value of other people when presented with propositional information about those people (e.g., seeing the phrase “hit a small child” presented with a picture of Sally) (Bliss-Moreau et al. 2008). In this example, the representation of “hit a small child” has an affective component which is integrated into the representation of Sally.
It is possible that some USs are exclusively experienced either via automatic, effortless associative processing or via effortful, controlled propositional processing (but not both), as Mitchell et al. and most dual-process theories suggest (for an extensive review, see Evans Reference Evans2008). A more likely scenario, however, is that the two types of processing are often active in parallel and serve to constrain each other to make meaning of a given stimulus in a given context. For example, the sound of gunfire is aversive and may have an automatic effect on the nervous system. But, for a person who has never experienced war, that automatic processing may be constrained by propositional information about the “shoot-'em-up” Western movies he or she remembers from childhood, resulting in a relatively neutral experience. For a war vet, the automatic processing may be constrained with propositional information gained in the experience of fighting and killing, resulting in a highly aversive experience. Propositional learning, even for a stimulus that has semantic meaning, is not required.
By focusing on how stimulus representations are changed as a result of internal experience, a whole host of learning phenomena can be united under one principle. Our hope is that by approaching learning from this perspective, the field will generate new hypotheses about the way that people learn about the world.
