All the co-authors of this commentary believe that associative learning theory is dead in the water. Penn and Povinelli (Reference Penn and Povinelli2007) have argued that associative learning alone is unable to account for causal reasoning in nonhuman or human animals. Hummel and Holyoak (Reference Hummel and Holyoak1997; Reference Hummel and Holyoak2003) have argued that associationist-style representations are inadequate for modeling human relational reasoning. Cheng (Reference Cheng1997) showed that both human and nonhuman causal learning involves rational inferences that go far beyond keeping track of covariation information. And Penn et al. (Reference Penn, Holyoak and Povinelli2008) have recently argued that animals of many taxa employ relationally structured, functionally compositional representations and that human minds, in particular, closely approximate the higher-order features of a classical “language of thought.” Indeed, one of us actually anticipated the thesis proposed by Mitchell et al. more than a decade ago: “Is it possible,” Waldmann and Holyoak (Reference Waldmann and Holyoak1992) asked in the final sentence of their paper, “that lower-order associative learning should be reduced to higher order causal induction, rather than vice versa?” (p. 235).
So, at first glance, it would seem that we would be strong supporters of the “propositional approach” proposed by this target article. We are not. Some cognitive processes in both human and nonhuman animals involve controlled, effortful inferences operating over highly structured relational representations, but Mitchell et al.'s claim that all learning is effortful, conscious, and propositional is unfounded and implausible.
Not all learning is conscious and inferential
Certainly, some learning involves hypothesis testing and conscious propositional beliefs – at least in humans. But there is overwhelming evidence that many forms of learning are implicit and non-inferential. To cite just the most obvious examples: priming, motor-skill learning, fear conditioning, and implicit category learning. Mitchell et al.'s thesis completely fails to account for these processes.
Mitchell et al. also miss the critical distinction between conscious awareness of a cognitive process and conscious awareness of the output of the process (e.g., Nisbett & Wilson Reference Nisbett and Wilson1977), a central feature of most dual-process models of cognition (Evans Reference Evans2008). Instead, Mitchell et al.'s thesis seems to resurrect Descartes' notion of a transparent mind fully conscious of its own reasoning processes.
We need more processes, not fewer
Like Mitchell et al., we are not fans of dual-process theories. But our problem is the opposite of Mitchell et al.'s: In our view, there are many more than two kinds of processes involved in human and nonhuman cognition. As Evans (Reference Evans2008) shows, the idea that there is one system that is purely conscious, inferential, and propositional, and one other that is purely automatic, implicit, and associative, is no longer sustainable. Mitchell et al.'s one-process theory fails to explain all the phenomena that motivated dual-process models to begin with (see Evans Reference Evans2008).
Propositions are not enough
We are big fans of propositions. But propositions alone are ill-suited for many aspects of biological cognition – for example, pattern matching and completion, graded semantic flexibility, and stimulus generalization. Any plausible model of biological cognition must incorporate both the structural features of propositional representations and the automatic, graded flexibility of distributed, associative representations (Hummel & Holyoak Reference Hummel and Holyoak1997).
Propositions are not just structured relations
Mitchell et al. briefly mention that propositions specify “the way in which events are related” (sect. 1, para. 5). While true, this is just the first step down a long path towards full-fledged propositions. There are many other critical features of propositions that Mitchell et al. omit, such as the capacity to systematically represent types, variables, roles, and higher-order relations, and to perform rule-governed operations over these representations in an inferentially coherent fashion (Hummel & Holyoak Reference Hummel and Holyoak1997).
Crucially, these propositional features do not form a package by nomological necessity (cf. Fodor & Pylyshyn Reference Fodor and Pylyshyn1988). In our view, non-human animals approximate certain features of propositions and not others (Penn et al. Reference Penn, Holyoak and Povinelli2008). By reducing propositions to structured relations, Mitchell et al. gloss over all the interesting computational and comparative challenges.
Causal learning is not monolithic
We agree that causal learning is not purely associative in either human or nonhuman animals (Cheng Reference Cheng1997; Penn & Povinelli Reference Penn and Povinelli2007; Waldmann & Holyoak Reference Waldmann and Holyoak1992). But this does not mean that all kinds of causal learning or all kinds of animals employ the same degree of propositional sophistication (Penn et al. Reference Penn, Holyoak and Povinelli2008). There is good evidence that nonhuman animals employ structured representations and are capable of first-order causal inferences. But this does not mean that rats employ the “higher-order reasoning processes” employed by humans (see Penn & Povinelli Reference Penn and Povinelli2007). In the case of Beckers et al. (Reference Beckers, Miller, De Houwer and Urushihara2006), for example, the rats' inferences can be modeled as a kind of sequential causal learning that does not require higher-order relational representations (Lu et al. Reference Lu, Rojas, Beckers, Yuille, Love, McRae and Sloutsky2008).
Rescorla-Wagner is not propositional
Mitchell et al. claim that the Rescorla-Wagner model can be thought of as a “simple mathematical model of propositional reasoning” (sect. 6.1, para. 10). Yes, the Rescorla-Wagner model could be implemented symbolically; but that does not make it a rational model. The Rescorla-Wagner model assumes a linear generating function and lacks a representation of causal power (Cheng Reference Cheng1997). Ironically, Mitchell et al. miss why associationist theories fail as rational models of causal reasoning.
PDP models are better and worse than Mitchell et al. claim
Mitchell et al. have an idiosyncratic view of parallel distributed processing (PDP) models. On the one hand, they claim that traditional PDP models can account for the propositional capabilities of humans. On the other hand, they claim that “a single node in a PDP model does not represent anything” (sect. 6.2, para. 3). They are wrong on both accounts. PDP models are incapable of representing the structured relations that Mitchell et al. claim are the sine qua non of learning (Hummel & Holyoak Reference Hummel and Holyoak1997). But this does not mean they represent nothing at all. Every node in a PDP network has some equivalence class, and this equivalence class is precisely what it represents. Just because this equivalence class does not correspond to something one can point at does not mean it does not exist.
Darwin and neuroscientists are not all wrong
Mitchell et al. admit that the neuroscientific evidence provides little support for their claim that all learning is propositional. But they dismiss this evidence as inconclusive. They take a similarly dismissive attitude towards the comparative evidence. They admit that Aplysia do not employ propositions. But they have no evolutionary explanation for what happened to the Aplysia's primordial associative learning mechanisms in more sophisticated creatures such as rats and humans. Did all this pre-propositional baggage simply shrivel up and die? We think not.
Many nonhuman animals have cognitive capabilities that go far beyond the automatic formation of simple links. But the degree to which propositional mechanisms are employed differs between different kinds of animals and between different kinds of cognitive process within a given individual. There are some ways in which humans are still very much like Aplysia. There are other ways in which we are unique.
Like Darwin, Mitchell et al. overestimate the propositional capabilities of nonhuman animals (Penn et al. Reference Penn, Holyoak and Povinelli2008). But worse, Mitchell et al. ignore the incremental and cumulative fashion in which evolution crafted the various kinds of minds on this planet. Darwin did not make that mistake.
