Memory and learning

We tend to interpret feedback in ways that confirm our pre-existing beliefs. This confirmation bias is often treated as irrational, but may have adaptive foundations. In this project, we propose a new Bayesian computational model of confirmation bias and a novel experimental paradigm to study its impact on learning. When faced with an ambiguous outcome, we must form the most accurate interpretation we can by making use of all available information, which includes our pre-existing beliefs. Confirmation bias may thus constitute an inductive bias that speeds up learning, analogous to missing data imputation. We test this theory using a reward learning task in which participants are only provided partial information about outcomes, allowing more leeway for subjective interpretation. We find that our Bayesian model better explains the dynamics of behavior and stated beliefs compared to more traditional learning models, supporting an adaptive basis for confirmation biased learning from repeated feedback.

Global matching is a key concept in many models of recognition memory which posits retrieval as a process of matching test probe against every stored memory representation to produce a measure of global similarity. However, to date many models have not adopted principled representations of words. While some memory models have advanced through adopting realistic semantic representations, little work has explored the consequence of integrating perceptual representation. A variety of orthographic representations of words has been proposed in the psycholinguistics literature to account for several orthographic similarity effects between word pairs, however, little contact has been made to recognition memory. The study aimed to firstly establish three key orthographic similarity effects in recognition memory, namely the replacement effect, exterior-letter effect and transposition effect, and secondly to compare four orthographic representations (i.e. slot-coding, closed-bigram, open-bigram, and the overlap model) in their ability to capture recognition memory data in a global matching framework. 162 participants completed a recognition memory study of words using unrelated lists where targets were paired with lures of different orthographic similarity types. Different orthographic representations were used to calculate a global similarity value for each test probe, which was then used to model recognition accuracy via Luce’s choice rule in a hierarchical Bayesian framework. Results showed clear replacement effect, adjacent and non-adjacent transposition effects and start-letter importance in recognition memory. Model selection results support the open-bigram coding being the best orthographic representation in recognition memory.

We built a hierarchical Bayesian model for the working memory updating task. This model jointly accounts for both responses and reaction times in the memory updating paradigm, which is a commonly used paradigm to measure working memory capacity. To model responses, we adopted a mutual interference framework from Oberauer & Kliegl (2006) that characterized activation levels of working memory items, and extended this framework into a Markov chain structure to characterize a wider range of responses. To model reaction times, we adopted a Wald diffusion framework where the Wald parameters were associated with activation levels of working memory items. This model allows us to investigate the mechanism underlying participant performance in the memory updating task under a joint theoretical framework. We applied this model to an empirical data set investigating the effects of working memory training. Modeling results revealed that training might not improve overall working memory capacity, but may lead to a general improvement in the speed of processing.

Existing models of memory posit separate processes for encoding and retrieval: the study of items is an endogenous process of item- and context-reinstatement, while retrieval occurs through an exogenous drift-diffusion procedure. We argue that the same iterative memory process underlying encoding also underlies recall and decision-making, and propose a new model of endogenous, context-based recall decisions. The simulated model explains documented empirical facts about accuracy and inter-response times (IRTs) in free-recall experiments. These facts include the distribution of IRTs, the increase in average IRTs over successive retrievals, and the negative relation between accuracy and IRTs. The model is isomorphic to a large class of drift-diffusion models, implying a memory-based microfoundation for these common decision models and their use in theories of free recall. We demonstrate the theory’s broad implications by applying it to more general decision problems.

Participants learned to classify a set of rock images into geologically-defined science categories. We then investigated the nature of their category-based memory representations by collecting old-new recognition data in a subsequent transfer phase. An exemplar model provided better qualitative accounts of the old-new recognition data than did a prototype or clustering model. However, to account for the variability in recognition probabilities among the old training items themselves, a hybrid-similarity exemplar model was needed that took account of distinctive features present in the items. The study is among the first to use computational models for making detailed quantitative predictions of old-new recognition probabilities for individual items embedded in complex, high-dimensional similarity spaces.