Backward blocking of relevance-indicating cues: Evidence for locally Bayesian learning

The phenomenon in associative learning called ‘blocking’ has been a central target for theories that aim to explain learning and attention. The training phases of the blocking procedure, described in detail below, can be run in backward order, yet still produce the blocking effect (e.g., Shanks,1985; Dickinson and Burke,1996; Kruschke and Blair, 2000). Explaining forward and backward blocking, along with other associative learning phenomena, has proven to be challenging. Some accounts of forward blocking include an attentional mechanism that modulates the influence of the cues on learning and/or responding (e.g., Kruschke, 2001; Mackintosh, 1975). Some accounts of backward blocking employ a Bayesian framework in which different combinations of associative weights are considered simultaneously, with more belief allocated to the combination that is most consistent with training items (e.g., Dayan and Kakade, 2001; Tenenbaum and Griffths, 2003). A theoretical framework that is able to combine the attentional and Bayesian approaches is called ‘locally Bayesian learning’ (Kruschke, 2006b). The framework is based on the idea that a learning system may consist of a sequence of subsystems in a feed-forward chain, each of which is a locally Bayesian learner. The argument for locally-learning layers was as follows. First, Bayesian learning is very attractive for explaining retrospective revaluation effects such as backward blocking, among many other phenomena (Chater, Tenenbaum, and Yuille, 2006). Second, globally Bayesian learning may also be unattractive for a number of reasons. In a large learning system, there are too many combinations of parameters to keep track of in a monolithic joint parameter space. Furthermore, many globally Bayesian models do not explain learning phenomena (such as ‘highlighting’, Kruschke, 2010) that depend on training order, because the models treat all training items as equally representative of