Investigating the influence of prior knowledge on future learning

HOW DOES THE PAST INFLUENCE WHAT WE LEARN IN THE FUTURE? INVESTIGATING THE ROLE OF PRIOR KNOWLEDGE ON LEARNING IN NEW SITUATIONS 

Summary

To navigate a complex world successfully, we need to gather knowledge about the rules that govern it. For this purpose, we continuously sample our surroundings and store important information in long-term memory, building an internal world model to predict future outcomes. These models, which represent regularities in the world, are known as schemas (Gilboa & Marlatte, 2017). It has been argued that the evolutionary purpose of remembering information is to anticipate the future (Schacter et al., 2007). Thus, memory is not just for storage, but guides our perception and allows proactive behavior, enabling us to generalize knowledge to situations we have never experienced.

Importantly, our expectations have a strong impact on how we process and encode information (van Kesteren et al., 2012). If we face a new situation, our predictions about what will happen guide our attention, thus allowing us to efficiently process and react to the incoming information. As a consequence, schemas make it easier to encode content that is in line with our convictions (Greve et al., 2019; van Kesteren et al., 2010), accelerating the integration of memories in the brain's stable long-term store (Tse et al., 2007, 2011; van Kesteren et al., 2010). That priorknowledge can accelerate the learning of new information has already been demonstrated for declarative memory tasks including associative learning (van Kesteren et al., 2013), sentence learning (van Kesteren et al., 2014), and melody encoding (Durrant et al., 2015). Recently, these findings have been extended to the procedural domain (King et al., 2019): when new movements were compatible with an already formed cognitive schema in a sequential motor task, new learning performance was enhanced.

Contrary to these findings, other results show that not only information that conforms to our expectations, but also events that critically violate our predictions are particularly memorable (Antony et al., 2021; Quent et al., 2022; van Kesteren et al., 2012). The impact of these prediction errors on learning has been studied extensively and modelled, e.g., in the reinforcement learning literature (Rescorla & Wagner, 1972; Sutton & Barto, 1998). While the role of schemas and surprising information on new memory encoding has been investigated in different lines of research, it is not well understood how these processes interact to allow flexible memory function. Moreover, little theoretical work has modeled this interaction of schematic and situation-unique information.

The aim of this project is to investigate how existing knowledge influences the encoding of new information in an interdisciplinary approach, combining experimental work (Schönauer) and computational modeling (Mehring). We will investigate how domain general schematic knowledge interacts with the new learning of specific experiences and how these new experiences in turn update existing knowledge structures. To these aims, we will vary both the strength of an existing schema and how well the to-be-learnt new content matches expectations derived from this schema. We will model this process using a Bayesian approach, where the strength of the existing schema is reflected in the width of the prior distribution and the sensory evidence of the to-be learnt information is entered as the likelihood. The prediction error between prior and likelihood will determine how much is learned from the new information and the posterior distribution will indicate the updated schema. We follow a Bayesian modelling approach as it has previously provided important insights in to the nature of human learning, for example in sensorimotor learning (Heald et al., 2021), concept learning (Lake et al., 2015), and conditioning (Gershman et al., 2017). Bayesian networks have also been proposed to model the influence of prior knowledge of task structure on facilitation of motor learning and generalization (Braun et al., 2010; Genewein et al., 2015; Lansdell & Kording, 2019) - a phenomenon that may have similar underlying mechanisms to the influence of schemas on learning investigated here. In contrast to other computational models focusing on the acquisition of concepts in the declarative domain (e.g., by explicitly modeling a hippocampal and a neocortical module, Singh et al., 2022), a Bayesian approach would allow us to study the influence of prior knowledge on learning of new information from a domain-general perspective and to identify learning mechanisms that guide learning in both procedural and declarative tasks. We could then not only model within-domain but also across domain influences of schema on new learning.

The main goals of this project are 1) to extend existing research on Bayesian updating to declarative memory, testing whether it can accurately model effects of schema on new learning performance, 2) to relate estimates of the Bayesian model reflecting the relevant schema, the sensory evidence, their mismatch, and schema updating to neural processes measured with fMRI and EEG, and 3) to test whether superior memory for schema congruent and schema incongruent information is mediated by different neural processes. We will first investigate how schemas guide new learning in the declarative memory domain, but plan to extend this work to motor learning, and their interaction (Schönauer & Mehring).