In a conditioning experiment, a light might signal food. There are at least two interesting questions about this conditional relationship. The first is a psychophysical question: Is the animal sensitive both to the events and to the conditional rule that describes the relationship between the events? The second question is: From what internal mechanism, or algorithm, does this conditional rule emerge? Mitchell et al. argue that learning emerges from an internal representation of the conditional rule as a proposition. This theory is contrasted with the view that the propositional representation emerges from internal associative links between the events that do not involve direct propositional mechanisms.
Cognitive penetrability and the absence of automaticity?
Because it rejects the automatic link processes in learning, a propositional mechanism is more parsimonious than “dual theories” that postulate two mechanisms. This parsimony is illusory because the target article authors themselves describe a second process involving memory retrieval. “Learning” is propositional, but memory is sometimes automatic. An association formed as a proposition may, in memory, become automatic. This distinction is important for Mitchell et al.'s argument since the propositional view relies on the cognitive penetrability (Brewer Reference Brewer, Weimer and Palermo1974) of conditioning experiments. That is to say, cognitive manipulations, such as instructions, can directly modulate conditioning without having direct experience with the conditioning events. Automatic memories would not be easily susceptible to cognitive manipulations. So, if a learned rule is penetrable, it is propositional; but, if the rule is not penetrable, it has become automatic and thus reflects memory, not learning. Cognitive penetrability as used here is therefore circular. If people are aware of the contingency, then instructions can influence behaviour; so the mechanism is propositional. If not, the measurement tool is not sensitive enough, or the behaviour has become automatic in memory.
Cognitive load manipulations designed to show that increasing load disrupts propositional reasoning provide empirical evidence for the presence of automatic links. Imposing a load, while participants are exposed to a redundant-cue blocking paradigm (A+ AX+), eliminates blocking because the load prevents the propositional process (De Houwer & Beckers Reference De Houwer and Beckers2003; Waldmann & Walker Reference Waldmann and Walker2005). However, even though the propositional process is challenged, participants still respond to both A and X. If the propositional process is blocked, then this must mean the learning here is “automatic.” Mitchell et al. imply that simple pairings are automatic memory processes and that the important issues in learning are phenomena like blocking and inhibition. Even associationists have argued that such phenomena are more susceptible to modulation than excitatory pairings (Bouton Reference Bouton2004; Swartzentruber & Rescorla Reference Swartzentruber and Rescorla1994). Mitchell et al. seem to be restricting their analyses to those learning phenomena that are least stable and most susceptible to modulation.
The autonomic conditioning preparations might not be ideal for demonstrating strong automatic learning. Most animal conditioning preparations generate learning in nearly 100% of animals. An important finding of many of the cognitive penetrability experiments is that some people learn, and some do not, but only those who “know” the contingency show conditioned responses. Hence, in many experiments, a significant proportion of participants do not learn. This hardly makes us confident that these are biologically prepared (Seligman Reference Seligman1970) learning preparations suitable for evaluating the formation of strong “automatic” associations. Penetrability may reflect weak conditioning, or perhaps no conditioning at all, and therefore begs the question as to what cognitive penetrability means for human conditioning. Furthermore, unlike the rats in a conditioning experiment involving shocks, participants in these autonomic learning tasks are not naïve. They have learned to expect motivational events following verbal instructions. They have learned to relax when told the teacher will not check their homework today. Is it surprising that they will feel anxious when told that a tone will be followed by a shock, or will relax when subsequently told it will no longer be followed by the shock?
We believe there are learned pairings that are not cognitively penetrable. These include evaluative conditioning (Baeyens et al. Reference Baeyens, Eelen and Van den Bergh1990a) and probably strong aversions and flavour preferences. Cultural dishes evoke strong positive reactions and these are learned. Few of us have a strong positive reaction to eating live grubs, but in some cultures some people do. In a classroom demonstration, we have difficulty getting our students to salivate at the thought of crunching a grub between their teeth, as we might with students within the appropriate culture. These reactions are learned but not penetrable. This conclusion would be again open to the circular criticism that these processes were overlearned and became automatic in memory.
Propositional learning as a valuable heuristic
We prefer lower-level psychological theories to explain rule learning and propositional reasoning, rather than have them as fundamental primitives of learning. If propositional learning is impossible to disconfirm and has not eliminated automatic learning, then is it valuable? We would argue that it is. High-level rule-based and symbolic cognitive views can generate important research. These cognitive views can either be a theory about internal mechanisms, or they can represent a sometimes normative description of the physical and statistical mechanisms in the world. For example, Tolman (Reference Tolman1948) challenged the S-R (stimulus-response) psychologists' automatic theories with his notion of the cognitive map, and this notion generated interesting research and challenged the automatic or associative approach. Now, however, animals' abilities to navigate in the world to goals are beginning to be explained by lower-level theories (Diez-Chamizo et al. Reference Diez-Chamizo, Sterio and Mackintosh1985).
Elsewhere, arguing against the notion that internal mental images of three-dimensional drawings may be rotated in much the same way they are in the physical world (a notion which generated a great deal of interesting research), Pylyshyn (Reference Pylyshyn1973) pointed out that this is not sustainable as a cognitive theory. Mental images are computations but not direct mental representations. Our own research on contingency learning in humans and rats was originally motivated by theories (e.g., Cheng Reference Cheng1997) that humans and animals internalized the notion of computing contingencies (Baker & Mackintosh Reference Baker and Mackintosh1979; Baker et al. Reference Baker, Murphy and Mehta2003). We have subsequently argued that for both animals and humans these computations of correlations, and even computations of the dependencies in multiple-event chains, emerge from a connectionist network (Baetu & Baker, in press; Murphy & Baker Reference Murphy and Baker2004; Wasserman et al. Reference Wasserman, Elek, Chatlosh and Baker1993).
Around the time of the “cognitive revolution,” a number of rule-based and symbolic computational models emerged. Interest in these models has diminished because connectionist models have accounted for much of the data they initially explained, as well as for some they did not (Shultz Reference Shultz2003). More recent symbolic models even incorporated sub-symbolic (automatic) modules (e.g., Anderson & Lebiere Reference Anderson and Lebiere1998). Moreover, it has long been known that connectionist models can generate truth tables, logical operations, and many other linear and nonlinear rules (McCulloch & Pitts Reference McCulloch and Pitts1943). “Higher-level” cognitive processes and developmental stages emerge from connectionist or automatic architectures (e.g., Shultz et al. Reference Shultz, Mareschal and Schmidt1994).
This brings us to propositional learning mechanisms. These propositions map the normative relationship between events in the world. Although we take issue with some of the data, we see Mitchell et al.'s work as a psychophysics of “propositional” relationships. They show the rules and mechanisms in the world people and animals can represent and pose a challenge for connectionist or other lower-level theories. We are confident that the principles of propositional learning will emerge from connectionist principles that, contrary to the authors' claims, provide an algorithm and not a computational description (Marr Reference Marr1982).
