Exploiting existing representations implies tapping an enormous domain, coextensive with human understanding and knowledge, and endowed with its own dynamics of piecewise and cumulative learning. The worth of Clark & Thornton's proposal depends on the relative importance of this dynamics and of the bottom-up mechanism they come to complement. Radical restructuring of theories and patterns of retrieval from long-term memory are discussed in the context of such an evaluation.

Connectionist networks excel at extracting statistical regularities but have trouble extracting higher-order relationships. Clark & Thornton suggest that a solution to this problem might come from Elman (1993), but I argue that the success of Elman's single recurrent network is illusory, and show that it cannot in fact represent abstract relationships that can be generalized to novel instances, undermining Clark & Thornton's key arguments.

Recoding the data for relational learning is both easier and more difficult than it might appear. Human beings routinely find the appropriate representation for a given problem because coding always takes place within the framework of a domain, theory, or background knowledge. How this can be achieved is still highly speculative, but should probably be investigated with hybrid models.

Clark & Thornton speculate that intervening in the real world might be a way of transforming type-2 problems into type-1, but they state that they are not aware of any definite cases. It is argued that the active construction of external representations often performs exactly this function, and that recoding via the real world is therefore common, if not ubiquitous.

Learning is the acquisition of knowledge, not of input/output mappings. The distinction between statistical and relational learning, as Clark & Thornton define those terms, is not useful because all human learning is relational. However, prior knowledge does influence later learning and the sequence in which learning tasks are encountered is indeed crucial. Simulations of sequence effects would be interesting.

Biological neural computation relies a great deal on architecture, which constrains the types of content that can be processed by distinct modules in the brain. Though artificial neural networks are useful tools and give insight, they cannot be relied upon yet to give definitive answers to problems in cognition. Knowledge re-use may be driven more by architectural inheritance than by epistemological drives.

Although Clark & Thornton's “trading spaces” hypothesis is supposed to require trading internal representation for computation, it is not used consistently in that fashion. Not only do some of the offered computation-saving strategies turn out to be nonrepresentational, others (e.g., cultural artifacts) are external representations. Hence, C&T's hypothesis is consistent with antirepresentationalism.

Clark & Thornton's conception finds an echo in implicit learning research, which shows that subjects may perform adaptively in complex structured situations through the use of simple statistical learning mechanisms. However, the authors fail to draw a distinction between, on the one hand, subjects' representations which emerge from type-1 learning mechanisms, and, on the other, their knowledge of the genuine abstract “recoding function” which defines a type-2 problem.

Clark & Thornton present the well-established principle that recoding inputs can make learning easier. A useful goal would be to make such recoding automatic. We discuss some ways in which incrementality and transfer in connectionist networks could attain this goal.

We argue that existing learning algorithms are often poorly equipped to solve problems involving a certain type of important and widespread regularity that we call “type-2 regularity.” The solution in these cases is to trade achieved representation against computational search. We investigate several ways in which such a trade-off may be pursued including simple incremental learning, modular connectionism, and the developmental hypothesis of “representational redescription.”

Clark & Thornton argue that the recoding functions which are used to solve type-2 problems are, at least in part, the ontogenetic products of general-purpose mechanisms. This commentary disputes this and suggests that recoding functions are adaptive specializations.

Self-deception is made unnecessarily puzzling by the assumption that it is an intrapersonal analog of ordinary interpersonal deception. In paradigmatic cases, interpersonal deception is intentional and involves some time at which the deceiver disbelieves what the deceived believes. The assumption that self-deception is intentional and that the self-deceiver believes that some proposition is true while also believing that it is false produces interesting conceptual puzzles, but it also produces a fundamentally mistaken view of the dynamics of self-deception. This target article challenges the assumption and presents an alternative view of the nature and etiology of self-deception. Drawing upon empirical studies of cognitive biases, it resolves familiar “paradoxes” about the dynamics of self- deception and the condition of being self-deceived. Conceptually sufficient conditions for self-deception are offered and putative empirical demonstrations of a kind of self-deception in which a subject believes that a proposition is true while also believing that it is false are criticized. Self-deception is neither irresolvably paradoxical nor mysterious, and it is explicable without the assistance of mental exotica. The key to understanding its dynamics is a proper appreciation of our capacity for acquiring and retaining motivationally biased beliefs.

“Self-deception” usually occurs when a false belief would be more rewarding than an objective belief in the short run, but less rewarding in the long run. Given hyperbolic discounting of delayed events, people will be motivated in their long-range interest to create self-enforcing rules for testing reality, and in their long-range interest to evade these rules. Self-deception, then, resembles interpersonal deception in being an evasion of rules, but differs in being a product of intertemporal conflict.

Mele's study of philosophical and psychological theories of self-deception informatively links the conceptual and dynamic aspects of self-deception and explicates it without positing mutually inconsistent beliefs, such as those occurring in two-person deception. It is argued, however, that he does not do full justice to the dissociation characteristic of self-deception and does not sufficiently distinguish self-deception from self-caused deception.

Mele views self-deception as belief sustained by motivationally biased treatment of evidence. This view overlooks something essential, for it does not reckon with the fact that in self-deception the truth is dangerously close at hand and must be repeatedly suppressed. Self-deception is not so much a matter of what one positively believes as what one manages not to think.

The views of self-deception that Mele attacks are thoroughly metaphorical, and should never have purported to imply the existence of real internal acts of deception. Research on self-deception, including Mele's appealing account, could be enriched and constrained by a broader investigation of the prevalent use of metaphor in thinking and talking about the mind.

How can self-deception avoid intention and conscious recognition? Nine processes of self-deception seem to involve biased links between plausible ideas. These processes allow self-deceivers to regard individual conclusions as fair and reasonable. Bias is only detected by comparing broad patterns, which individual self-deceivers will not do.

This commentary defends intentionalist accounts of self-deception against Mele by arguing that: (1) viewing self-deception on the model of other-deception is not as paradoxical as Mele makes out; (2) the paradoxes are not entailed by the view that self-deception is intentional; and (3) there are two problems for Mele's theory that only an intentionalist theory can solve.

Mele's analysis of self-deception is persuasive but it might also be useful to consider the varieties of self-deception that occur in real-world settings. Instances of self-deception can be classified along three dimensions: implicit versus explicit, motivated versus process-based, and public versus private. All three types of self-deception have implications for the scientific research enterprise.

The simultaneous possession of conflicting beliefs is both possible and logical within current models of human cognition. Specifically, evidence of lateral inhibition and state-dependent memory suggests a means by which conflicting beliefs can coexist without requiring “mental exotica.” We suggest that paradoxical self-deception enables the self-deceiver to store important information for use at a later time.