Episodic memory

People often think about counterfactual possibilities to an event and imagine how it could have been otherwise. The study of how this occurs is central to many areas of cognitive psychology, including decision making, social cognition, and causal judgment; however, cognitive models of the memory processes at play during the generation of counterfactual thoughts have not yet been developed. Inspired by theories of list recall and semantic memory search, we build a formal model that examines how a sequence of counterfactual thoughts is retrieved from a set of all possible counterfactuals. Our approach takes the form of a Markov random walk over items in memory and allows for the activation of a counterfactual item to depend on its desirability, probability of selection, language frequency, and semantic similarity with the previously retrieved item. In this way, our model parametrically instantiates prior theories of counterfactual generation within a statistical model that can be fit to data from counterfactual generation tasks. Across three experiments, we show that our model describes and predicts the sequence of counterfactual thoughts that come to mind in response to a particular event, as well as the effects of these counterfactuals on subsequent evaluations and decisions. Our model can also explain key qualitative patterns in counterfactual generation and model the effects of contextual variables such as priming. Overall, our work shows how existing theories of counterfactual generation can be combined with quantitative models of memory search to provide new insights about the generation and consequences of counterfactual thinking.

Alzheimer’s disease leads to a decline in both episodic and semantic memory. Free recall tasks are commonly used in assessments designed to diagnose and monitor cognitive impairment, but tend to focus on episodic memory. Our goal is to understand the influence of semantic memory on the sequence of free recall in a clinical data set. We develop a cognitive process model that allows for the influence of semantic similarity and other stimulus properties on the order of free recall. The model also incorporates a decision process based on the Luce choice rule, allowing for different levels of response determinism. We apply the model to a real-world data set including free recall data from 2392 Alzheimer’s patients and their caregivers. We find that semantic similarity between items strongly influences the order of free recall, regardless of impairment. We also observe a trend for response determinism to decrease as impairment increases.

Memory reactivation can be observed during sleep or wakefulness in human and rodent brains, and is believed to be crucial for memory consolidation (Lewis and Bendor, 2019). A similar strategy, namely rehearsal or replay, is proven to be effective in mitigating, or even overcoming the catastrophic forgetting problem in neural network (NN) modelling and applications (Robins, 1995; Kumaran and McClelland, 2012). Generative replay (GR) (van de Ven, Siegelmann and Tolias, 2020) and experience replay (ER) (Káli and Dayan, 2004) are the two common replay strategies. While GR produces replay samples from random activations in a generative NN, ER revisits exact copies of past training samples preserved in memory storage. Although ER (without memory limits) yields better results and is thus deployed more in applications (e.g., machine learning), GR is computationally more efficient and biologically more plausible. In this study we chose restricted Boltzmann machines (RBMs) as our primary NN model. In addition to ER and GR, we consider a new strategy cued generative replay (cGR), which uses replay cues that are partially correct activations rather than completely random activations in standard GR. We propose two indices, evenness and exactness to measure the quality of replay samples. GR, in contrast to ER, yielded more balanced but less accurate replay (high evenness, low exactness), but their performance was largely dependent on the replay amount. We found that cGR could outperform both by improving replay quality.

Recognition memory for short lists shows effects of recency and primacy, which are typically explained using different mechanisms. I propose a single account that jointly explains recency and primacy. This account uses the same mechanisms that explain the dynamics of encoding and recognition of associations between items (Cox & Criss, 2020). When two items are presented simultaneously, separate representations of those items are gradually built by sampling perceptual and semantic features (Cox & Shiffrin, 2017). As this is happening, associative features are formed by making conjunctions between item features, and these associative features then end up shared between the originally separate item representations. I propose that this same process operates when items are presented in sequence, except that instead of forming associative features between two items, associative features conjoin features of a single item with those of its temporal context, which consists of features from preceding items (e.g., Howard & Kahana, 2002; Logan, 2020). This model accounts for recognition accuracy and response times across list lengths and serial positions (Nosofsky, Cox, Cao, & Shiffrin, 2014) as well as facilitation when study order is preserved at test (Schwartz et al., 2005). Recency and facilitation occur because the associative features between the test item and the prevailing temporal context are a strong match to the associative features formed at study. Primacy occurs because temporal contexts are stored in memory and they tend to over-represent early items. The same mechanisms that form associations between items within a trial can also explain associations between trials.

Recent qualitative reviews show that sleep-dependent memory consolidation (SDMC) effects are highly task dependent. A growing body of research argues that encoding-related spontaneous reactivation and reactivation due to memory cueing during sleep play a causal role in SDMC, specifically for associative information and gist abstraction (Lewis, Knoblich & Poe, 2018). To better understand the relationship between task-dependency, reactivation, and rapid generalization a formal framework is necessary. We argue that an exemplar-based framework (Hintzman, 1986) is complementary to the existing connectionist computational models of reactivation (e.g: Kumaran & McClelland, 2012). By modelling offline reactivation as internally generated cued recall we can account for numerous behavioural SDMC findings (including episodic inference tasks, categorization, motor memory), some of which have been shown to be related to SWS. We discuss predictions regarding the effects of interference, memory strength, context and how they relate to existing verbal theories of SDMC. We conclude that recurrent similarity-based generalization is an ideal algorithm for modelling consolidation of newly encoded memories.