Mitchell et al. have made a number of important contributions to our understanding of human contingency learning. They cite many compelling studies that demonstrate the strong influences of verbal instruction and processing resources on learning. Along the way they have become committed to a particular representational format that may limit their ability to explain the breadth of learned human behavior.
We have two main concerns with their proposal that learning is the result of controlled reasoning processes that operate on propositional representations. First, it is unclear that many of the empirical findings they describe automatically follow from a propositional approach. These would instead seem to depend on the particular models of learning and memory that one employs, not on the use of a propositional format per se. Second, if Mitchell et al. are indeed arguing that all learned knowledge is encoded as propositions with attached levels of belief – a very strong claim – they need to provide a convincing account of how we could learn to perform complex, context-sensitive behaviors using processes that operate only on propositions and beliefs. Without such an explanation, their theory can only speak to a restricted set of data drawn from standard human contingency learning paradigms, and not to human and animal learning in its broader sense. We elaborate these points below.
In their discussion of cognitive load, Mitchell et al. describe a series of studies by De Houwer and Beckers (Reference De Houwer and Beckers2003) who found a reduction of forward blocking in a standard contingency learning task when a moderately engaging secondary task was introduced. At the same time, participants continued to demonstrate robust learning of the explicit contingencies they experienced during training. Although Mitchell et al. do not provide an explanation as to why blocking should require more cognitive resources than learning all of the other cue-outcome relationships in the task, we have been interested in whether the ability to retrieve relevant memories may play a role in generating these effects.
We have recently explored cue-competition effects in two tasks using common stimuli and abstract contingencies but different task demands: a self-paced prediction task that is similar to a standard contingency learning experiment, and a fast-paced, cued reaction task (Sternberg & McClelland, in preparation). In both tasks, participants saw one or two cue objects on each trial, some of which were usually followed by the appearance of a dot. In the prediction task, participants had unlimited time to observe the cue objects before pressing a key to predict whether the dot would appear on that trial. After making their prediction, the actual outcome appeared and feedback was given. In the reaction task, on trials where the dot occurred, it appeared 350 msec after the cue object(s), and participants had to press a key within 275–400 msec (this deadline decreased at a constant rate during training). On trials where the dot did not appear, participants had to refrain from responding. Both groups were able to learn the explicit contingencies they experienced during training, as revealed by reaction times (RTs) and contingency ratings in the reaction task, and by test-trial predictions and contingency ratings for the prediction task. However, while prediction participants showed clear competition effects, evidence for competitions effect in the reaction task was scant.
Mitchell et al.'s view that knowledge of the contingencies is stored in the form of propositions does not, in our view, shed light on the differences between these conditions. The form of storage could be propositions, images, or something else – regardless, the explanation could hinge on constraints on the time available for retrieval of relevant prior episodes from memory. On this view, blocking would depend on considering both the relevant singleton and the related pair at the same time, even though these events are not presented together, whereas learning the direct events would depend only on recalling past experiences with the presented stimulus. Because the reaction task reduces the time available for recollection, we would expect to see a reduction in cue-competition effects in this task compared to those observed in the prediction task.
We also consider it to be uncontroversial that the operation of any learning process should depend on the availability of cognitive resources such as attention and memory. There have been many theories of multiple interacting learning processes that make no claims about a completely autonomous and resource-free associative learning system (e.g., McClelland et al. Reference McClelland, McNaughton and O'Reilly1995). As it is likely that a number of interacting neural systems can spring into action when a novel or familiar stimulus is encountered, it will be important to continue to study in more detail the role of attention (see Kruschke & Blair Reference Kruschke and Blair2000; Kruschke et al. Reference Kruschke, Kappenman and Hetrick2005) and memory (McClelland & Thompson Reference McClelland and Thompson2007; Vandorpe et al. Reference Vandorpe, De and Beckers2007b) in learning and reasoning.
The propositional account that Mitchell et al. advocate also may run into difficulties addressing details of situation-specific response behaviors. The same contingency may be learned in different behavioral contexts, and require different kinds of responses, and the responses may not transfer even if the contingencies remain the same. For example, we would not expect that participants who had learned cue-outcome contingencies in our prediction task would immediately show fast RTs to relevant items if tested using our reaction task, even if they were verbally instructed that the same contingencies they had previously learned still held. It seems likely that something more like a sensory-motor skill of responding to particular stimuli has been acquired by these participants. Such skills, and not simply verbal propositions, seem fairly clearly to underlie abilities like playing an instrument or driving a car.
Finally, Mitchell et al.'s current theory sheds little light on the process of learning complex behaviors. Consider Cleeremans and McClelland (Reference Cleeremans and McClelland1991), who trained participants to make button responses to sequences of visual stimuli that were generated by a moderately complex finite state grammar. They found that participants show graded sensitivity to the conditional probabilities between items, and that this sensitivity developed gradually across training sessions. They were able to capture the learning trajectories in the human data using a simple recurrent network (SRN) model. While the structure of the information in the task could certainly be represented in a propositional form (as it was represented in this way in the experiment script), the SRN provided an account of the process of learning through time, in addition to its final state. The key ingredient here appears to be some sort of gradual strengthening process, rather than the presence or absence of a propositional statement of contingency.
Mitchell et al. conclude by presenting us with a tall order. Because they find little evidence for an automatic link-formation system, they suggest that we should recast all of the existing literature on human learning as evidence about the operation of a single propositional learning system. This conclusion seems to us to present a false dichotomy. The studies the authors present on cognitive load and verbal instruction manipulations are indeed important and useful challenges to simple associative accounts of contingency learning, but by embedding all knowledge in a propositional form, Mitchell et al. may sacrifice an account of how we learn complex behaviors.
