Rumbaugh et al. (Reference Rumbaugh, King, Beran, Washburn and Gould2007) have argued that stimulus-response learning and reinforcement as constructs, as put forth in classic behaviorism over the course of the past century, are in need of redefinition. The essence of the redefinition entails contiguous stimuli sharing their response-eliciting properties as a function of their relative attributes and strengths (e.g., their saliences), forming amalgams. These amalgams are not simply linkages between stimuli; they are new entities based on the properties of the stimuli, their response-eliciting properties, and any other salient features of the environment. The ease with which amalgams form, is a function of the natural history of the species and the constructive bias of its neural system.

The neural system integrates the amalgams into templates that metaphorically define a knowledge base, which includes information about the individual and its assessed capabilities in relation to the resources and risks of its ecological niche. From this knowledge base, emergent behaviors and capacities, with no history of specific training, might take form as overarching principles and rules to service adaptation rationally and creatively in both familiar and novel challenges. This theoretical stance has much in common with that of Mitchell et al. Neither theory supports the idea of reinforcement as a direct line to the CR (conditioned response); nor does either theory support the notion of simple S-S (stimulus-stimulus) links as part of a dual-learning system.

This salience theory is silent on the issue of consciousness; however, its provision for emergent behaviors and capabilities allows for consciousness to emerge as a product of the complexity of the nervous system and its knowledge base. Consciousness, once functioning, might selectively interact with the systems of templates that are always subject to modification across time and experience. Thus, consciousness and symbol-based logic (and propositional knowledge) might either monitor or modify systems of templates and the amalgams they organize. Mitchell et al. make a good case against unconscious link-based learning systems, but they do admit the existence of some evidence for a learning system in which subjects react contrarily to their stated beliefs. Mitchell et al. correctly suggest these findings could be the result of an “imperfect” propositional system, where the subjects have created beliefs about certain contingencies, but those contingencies are flawed in some way. However, there is no reason to assume that all learning, even if propositional in nature, must be fully conscious and accessible.

To illustrate the viability of the salience theory in these situations, we turn to some of Mitchell et al.'s examples. Mitchell et al. discuss the concept of blocking (sect. 4.2), where a pre-trained CS (A+) blocks the association of a new CS (T+) and the US (AT+). This is explainable under the salience theory, since the pre-trained CS (A) would be highly salient and would overshadow any new competitive amalgam formation unless the new CS (T) were equally or more salient. Additionally, the finding that embedded tasks (cognitively demanding tasks presented simultaneously with the presentation of the AT+ condition) result in subjects' better learning the T condition (sect. 4.2), is supported by salience theory. In this case, the embedded/distracter task reduces salience of A during the presentation of AT+, increasing the salience of T and resulting in some learning of the T+ contingency. Finally, in the case when AT+ was trained, how did the presentation of A− serve to increase the response to T? Perhaps A and T had become equally salient constituents within new amalgams, but subsequent A− training reduced the saliency of A, in effect, increasing the relative saliency of T (sect. 4.3).

Both Mitchell et al.'s and Rumbaugh's theories are equally supported by findings from a long history of nonhuman cognitive studies, some of which were mentioned in the target article. One of the most reliable indications of learning abilities in nonhuman primates not mentioned in the target article is the transfer index (Rumbaugh & Washburn Reference Rumbaugh and Washburn2003). This index shows the ability of individuals trained in a simple one-choice test (choose between A and B) to switch a preference from a trained positive stimulus (A+) to a new positive stimulus (B+), given a certain level of ability on the first trained stimulus (67% or 84%). In other words, when your performance on the first task (always choose A) is increased from 67% to 84%, how is your performance on the second task (switch to always chose B) affected? These studies show that smaller-brained primates are adversely affected by augmented learning; that is, the more they learn in the first task, the worse they do on the second. In larger-brained primates, this finding is exactly reversed (the correlation between transfer index performance and various measures of brain size range from .79 to .82). Additionally, transfer index performance is strongly affected by early environmental conditions (see Rumbaugh & Washburn Reference Rumbaugh and Washburn2003, for a review). These results are difficult to explain with a straight link-based learning system and suggest that learning processes are altered by elaboration of the brain and by rearing.

Further evidence for a propositional/salience-based learning system comes from language learning in chimpanzees and bonobos. These studies include findings in rapid symbolic associations (association of symbols to new referents on a one-trial basis; Lyn & Savage-Rumbaugh Reference Lyn and Savage-Rumbaugh2000) and “representational play” – treating a toy or toy stand-in as if it were something else (Lyn et al. Reference Lyn, Greenfield and Savage-Rumbaugh2006). Moreover, complex mental representations of symbols have been documented by detailing the errors in a vocabulary test (Lyn Reference Lyn2007). These errors include choosing a visual symbol (e.g., a lexigram that represents “key”) based on a photograph of a referent that has an auditory similarity (“TV”) – indicating that all levels of representation are activated in the choice task. Additionally, emergent behavior (such as the initial language learning in a bonobo; Savage-Rumbaugh et al. Reference Savage-Rumbaugh, McDonald, Sevcik, Hopkins and Rupert1986) frequently appears during cognitive training in nonhumans, indicating that nonhumans are able to respond not just based on their learned contingencies, but rather by construction of an emergent, unlearned (self-generated) contingency, a feat difficult to explain through standard S-R (stimulus-response) link-based learning.