Year: 2023
Type: dissertation
Abstract: Reactivation and replay in both biological and artificial agents offer significant computational \nadvantages. Research has demonstrated that these processes can lead to \naccelerated learning, reduced forgetting, and the reorganization or augmentation of \nexperiences, further supporting planning and generalization. These reactivations can \ntranspire during both online and offline periods. Specifically, online reactivation pertains \nto the immediate reactivation of neural activity patterns during wakefulness. \nIn contrast, offline reactivation takes place during rest or sleep intervals, when the \nbrain is disengaged from external tasks. In these moments, the brain revisits and \nreactivates neural activity patterns initially established during wakeful states. \nA significant gap exists in our understanding however of how the brain determines \nwhich information to reactivate during its limited offline periods. Specifically, the conditions \nunder which offline reactivation contributes to adaptive generalization remain \nambiguous, especially as recent reviews have highlighted the limitations of the extent \nof offline benefits (Cordi & Rasch, 2021; Lerner & Gluck, 2019a). This uncertainty is \nnot trivial; given the pivotal role offline periods play in memory consolidation, comprehending \nthe mechanisms underlying adaptive consolidation and generalization is \nparamount. Such understanding can offer insights into optimizing learning strategies \nand determining the best times to take breaks. Furthermore, it can guide the \ndevelopment of artificial agents designed for continuous learning and might on day \nserve as reliable assistants in our daily activities. Additionally, a deeper grasp of \nhealthy memory consolidation processes can pave the way for identifying early indicators \nof consolidation breakdowns, whether due to aging, mental health conditions, \nor other factors. This knowledge could also usher in innovative applications, such as \nfacilitating learning or unlearning during rest periods, where memory might be more \nmalleable than during wake. \nThe central aim of this dissertation is to investigate the merit of a model-based \nanalysis of offline reactivation-dependent consolidation following episodic learning. \nThis exploration seeks to understand the potential mechanistic contributions of hippocampal \noffline reactivations to the generalization observed in animals when engaged \nin episodic and serial learning tasks. Another pivotal objective is to identify moderating \nvariables that can account for why extended post-learning retention intervals, \nwhich encompass “offline reactivations,” sometimes result in noticeable generalization \nbenefits, while in other instances they either have no effect or even lead to decreased \nperformance upon delayed retrieval. Augmenting this empirical research with a quantitative \nmeta-analysis can further validate claims regarding these moderator variables. \nWith a deeper understanding of these conditions, it is anticipated that advancements \nin pattern analysis of generalization, combined with neuroimaging techniques like \nfMRI, can be employed to pinpoint not only the “loci” of generalization but also the \nspecific nature and evolution of generalization over time. \nThis dissertation, at its core, encapsulates my journey in learning diverse methodologies \nand approaches with the overarching aim of producing research that is both \nreproducible and replicable. The significance of this work extends beyond its immediate \nfindings, offering broader implications for the field at large. The cognitive \nmodeling of ostensibly simple tasks, such as transitive inference, holds promise. It \nnot only facilitates precise communication through mathematical paradigms among \nresearchers from disparate disciplines but also engenders interdisciplinary adversarial \ncollaborations. Such collaborations can catalyze the design of experiments situated \nat the intersection of contending formal theories, thereby fostering incremental advancements \nin the field. The adoption of online experimentation, particularly within \nthe domain of time and sleep-dependent consolidation, represents a relatively nascent \napproach. Our endeavors in this realm are anticipated to pave the way for future \nstudies, characterized by enhanced statistical robustness. In tandem with this, our \nconcise meta-analysis of the extant datasets serves as a precursor to more expansive \nmeta-analytical endeavors, poised to elucidate the moderators of generalization and \nprovide direction for subsequent research. Furthermore, this dissertation endeavors \nto illuminate an efficacious methodology for examining representational shifts over \ntime. This is achieved through a within-subject multi-session design that combines \nremote learning conditions with in-scanner retrieval, employing a localizer task to \nscrutinize the representational geometry underpinning inference. While each of these \nmethodological innovations might not be unprecedented in isolation, their confluence \nwithin this research area is rare. Such a synthesis holds the potential to augment the \ncurrent state of sleep and memory research. \nHowever, it’s important to note certain limitations that might influence the interpretation \nor generalizability of the findings. The scope of this dissertation is defined \nby its concentrated focus on theories of generalization discussed in Chapter 1, particularly \nemphasizing reactivation as the primary underlying mechanism. In terms of \ntasks, the research is primarily centered on the phenomena of associative inference, \nwith a specific emphasis on transitive inference throughout. Additionally, while the \nmain emphasis is on offline consolidation, this research does not include any direct \nphysiological measures of reactivation during the offline period. The analysis is anchored \nsolely in recall performance, with only minimal measures following immediate \nrecall. Regrettably, none of the experiments undertaken involve measuring immediate \npost-learning rest or sleep physiology, which could provide deeper insights into the \nprocess of offline consolidation. \nFor this research, a multifaceted methodological approach was adopted to delve \ninto the intricacies of reactivation-dependent generalization. A vector-based memory \nmodels were crafted using both MATLAB and Python, facilitating the simulation of \nspecific experiments. These simulations were useful in shedding light on reactivationdependent \ngeneralization, resonating with the overarching goals of this dissertation. \nOn the behavioral analysis front, the methodology predominantly hinges on logistic \nmixed-models and employs remote web-based experimentation to dissect timedependent \nconsolidation. Furthermore, univariate random-effects models have been \nemployed for a meta-analysis of the transitive inference findings, as well as a metaregression \nanalysis of moderator variables. To enhance the depth of the research, \na follow-up experiment was analyzed using model-based representational similarity \nanalysis, complemented by fMRI data. \nThe dissertation is structured to provide a rigorous exploration of the topic at \nhand. Chapter 1 commences with a theoretical exposition, delineating both the \nclassical and more recent paradigms that inform our understanding of human generalization. \nAs the discourse advances, attention is directed towards cognitive models \nof generalization, with an emphasis on those that can produce offline generalization \nphenomena. In Chapter 2, a methodical comparative analysis is undertaken, juxtaposing \ntwo salient cognitive models: REMERGE and MINERVA2. This chapter \nunderscores the efficacy of model-based methodologies in the study of generalization, \npositing MINERVA2 as an exemplar baseline model. Its value lies in its capacity to account \nfor an array of time-dependent findings with a parsimonious set of parameters. \nChapter 3 presents a triad of empirical investigations centered on the details of time \nand sleep-dependent generalization as they manifest following the transitive inference \ntask. Of these, one is a reanalysis of an extant dataset, while the subsequent two \ncollected as part of this dissertation, adopt divergent design paradigms: the former \nadhering to the conventional between-subject design and the latter a novel withinsubject \napproach. Transitioning to Chapter 4, the narrative engages in a reevaluation \nof two secondary datasets published in the realm of transitive inference. This chapter \nends with a meta-analytical synthesis, amalgamating insights from our three primary \nexperiments, the reinterpreted datasets, and additional published research, thereby \nfacilitating a comprehensive examination of time and sleep-dependent effects. Concluding \nthe dissertation, Chapter 5 introduces the final empirical endeavor, which \nharnesses neuroimaging techniques to probe the representational geometry underpinning \nsuccessful inference at delayed test, building upon the previously piloted \nwithin-subject study of transitive inference.
Author: Foldes, Tamas
Institution: Cardiff University
Language: English
Topic: Memory and Neural Mechanisms
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W6992894824
