Jeremy R. Manning

What is a second recounting of an experience a recounting of? Memories change with time and with retrieval, so each recounting may be shaped less by the experience itself than by how we last recounted it. Here we use a geometric framework to compare two recountings of the same experience. We model experiences and recountings as trajectories through a semantic embedding space, then segment each trajectory into discrete events. We apply our approach to data collected from participants who watched a television episode, verbally recounted it immediately afterward, then recounted it again one week later. We find that the trajectories of participants' delayed recountings resemble their own immediate recountings more closely than they resemble the episode itself. We also find that this resemblance is person-specific enough to identify one of an individual's recountings of the experience from the geometry of their other recountings. The delayed recountings' trajectories were also well described as geometrically simplified versions of immediate recountings' trajectories, in that the immediately recounted events that least contributed to a trajectory's shape were the events most likely to be omitted from the delayed recounting. These findings suggest that the act of retrieving a memory after a delay draws on a person-specific distillation of the experience, rather than the original experience as it was first encountered. Further, the distillation itself appears to proceed in a way that best preserves the shape of the original recounting's trajectory.

Real-world conceptual knowledge is complex, multifaceted, and substantially over-simplified in most laboratory studies. Here we develop a mathematical framework, based on natural language processing models, for tracking and characterizing the acquisition of real-world conceptual knowledge. Our approach embeds each concept in a high-dimensional representation space where nearby coordinates reflect similar or related concepts. We test our approach using behavioral data from participants who answered small sets of multiple-choice quiz questions interleaved between watching two Khan Academy course videos. We apply our framework to the videos’ transcripts and the text of the quiz questions to quantify the content of each moment of video and each quiz question. We use these embeddings, along with participants’ quiz responses, to track how the learners’ knowledge changed after watching each video and predict their success on individual quiz questions. Our findings show how a small set of quiz questions may be used to obtain rich and meaningful insights into what each learner knows, and how their knowledge changes over time as they learn.

We show that large language models (LLMs) can be used to distinguish the writings of different authors. Specifically, an individual GPT-2 model, trained from scratch on the works of one author, will predict held-out text from that author more accurately than held-out text from other authors. We suggest that, in this way, a model trained on one author's works embodies the unique writing style of that author. We first demonstrate our approach on books written by eight different (known) authors. We also use this approach to confirm R. P. Thompson's authorship of the well-studied 15th book of the Oz series, originally attributed to F. L. Baum.

We develop a mathematical framework, based on natural language processing models, for tracking and characterizing the acquisition of conceptual knowledge. Our approach embeds each concept in a high-dimensional representation space, where nearby coordinates reflect similar or related concepts. We test our approach using behavioral data from participants who answered small sets of multiple-choice quiz questions, interleaved between watching two course videos from the Khan Academy platform. We apply our framework to the videos' transcripts and the text of the quiz questions to quantify the content of each moment of video and each quiz question. We use these embeddings, along with participants' quiz responses, to track how the learners' knowledge changed after watching each video. Our findings show how a small set of quiz questions may be used to obtain rich and meaningful high-resolution insights into what each learner knows, and how their knowledge changes over time as they learn.

Unlike temporally symmetric inferences about simple sequences, inferences about our own lives are asymmetric: we are better able to infer the past than the future, since we remember our past but not our future. Here we explore whether there are asymmetries in inferences about the unobserved pasts and futures of other people’s lives. In two experiments (analyses of the replication experiment were pre-registered), our participants view segments of two character-driven television dramas and write out what they think happens just before or after each just-watched segment. Participants are better at inferring unseen past (versus future) events. This asymmetry is driven by participants’ reliance on characters’ conversational references in the narrative, which tend to favor the past. This tendency is also replicated in a large-scale analysis of conversational references in natural conversations. Our work reveals a temporal asymmetry in how observations of other people’s behaviors can inform inferences about the past and future.

To efficiently yet reliably represent and process information, our brains need to produce information-rich signals that differentiate between moments or cognitive states, while also being robust to noise or corruption. For many, though not all, natural systems, these two properties are often inversely related: More information-rich signals are less robust, and vice versa. Here, we examined how these properties change with ongoing cognitive demands. To this end, we applied dimensionality reduction algorithms and pattern classifiers to functional neuroimaging data collected as participants listened to a story, temporally scrambled versions of the story, or underwent a resting state scanning session. We considered two primary aspects of the neural data recorded in these different experimental conditions. First, we treated the maximum achievable decoding accuracy across participants as an indicator of the “informativeness” of the recorded patterns. Second, we treated the number of features (components) required to achieve a threshold decoding accuracy as a proxy for the “compressibility” of the neural patterns (where fewer components indicate greater compression). Overall, we found that the peak decoding accuracy (achievable without restricting the numbers of features) was highest in the intact (unscrambled) story listening condition. However, the number of features required to achieve comparable classification accuracy was also lowest in the intact story listening condition. Taken together, our work suggests that our brain networks flexibly reconfigure according to ongoing task demands and that the activity patterns associated with higher-order cognition and high engagement are both more informative and more compressible than the activity patterns associated with lower-order tasks and lower engagement.

Our memories of episodic experiences are represented and organized by contexts. Events that share overlapping contextual features such as spatio-temporal proximity are more likely to be organized together in our memories. But how do we represent and remember other people? The predominant view in social cognition is that people are represented as a bundle of attributes such as personality traits. However, in real-world contexts we most often learn about others by observing or participating in social interactions, which inform our understanding of interpersonal-relationships. In this study, we explore how people spontaneously represent and remember others after watching a character-driven television show. We find that memory for individual characters is organized around interpersonal relationships rather than person- or situation-specific attributes (i.e., actions, traits, social groups, or locations). Moreover, even the subtle order in which participants recalled the names of each character can be predicted by graphs of that character’s social network. These findings suggest that the way we naturally represent and remember people is largely driven by their connections with others. We believe our work provides new avenues of inquiry that move beyond person-centric approaches in social cognition and a novel perspective on contextual features for memory research.

Brain activity patterns are highly flexible and often complex, but also highly structured. Here we examined how fundamental properties of brain activity patterns relate to ongoing cognitive processes. To this end, we applied dimensionality reduction algorithms and pattern classifiers to functional neuroimaging data collected as participants listened to a story, temporally scrambled versions of the story, or underwent a resting state scanning session. These experimental conditions were intended to require di ff erent depths of processing and inspire di ff erent levels of cognitive engagement. We considered two primary aspects of the data. First, we treated the maximum achievable decoding accuracy across participants as an indicator of the “informativeness” of the recorded patterns. Second, we treated the number of features (components) required to achieve a threshold decoding accuracy as a proxy for the “compressibility” of the neural patterns (where fewer components indicate greater compression). Overall, we found that the peak decoding accuracy (achievable without restricting the numbers of features) was highest in the intact (unscrambled) story listening condition. However, the number of features required to achieve comparable classification accuracy was also lowest in the intact story listening condition. Taken together, our work suggests that our brain networks flexibly reconfigure according to ongoing task demands, and that the activity patterns associated with higher-order cognition and high engagement are both more informative and more compressible than the activity patterns associated with lower-order tasks and lower levels of engagement. Significance Statement: How our brains respond to ongoing experiences depends on what we are doing and thinking about, among other factors. To study how brain activity reflects ongoing cognition, we examined two fundamental aspects of brain activity under di ff erent cognitive circumstances: informativeness and compressibility. Informativeness refers to the extent to which brain patterns are both temporally specific and consistent across di ff erent people. Compressibility refers to how robust the informativeness of brain patterns is to dimensionality reduction. Brain activity evoked by higher-level cognitive tasks are both more informative and more compressible than activity evoked by lower-level tasks. Our findings suggest that our brains flexibly reconfigure themselves to optimize di ff erent aspects of how they function according to ongoing cognitive demands.

We perceive, interpret, and remember ongoing experiences through the lens of our prior experiences. Inferring that we are in one type of situation versus another can lead us to interpret the same physical experience differently. In turn, this can affect how we focus our attention, form expectations about what will happen next, remember what is happening now, draw on our prior related experiences, and so on. To study these phenomena, we asked participants to perform simple word list-learning tasks. Across different experimental conditions, we held the set of to-be-learned words constant, but we manipulated how incidental visual features changed across words and lists, along with the orders in which the words were studied. We found that these manipulations affected not only how the participants recalled the manipulated lists, but also how they recalled later (randomly ordered) lists. Our work shows how structure in our ongoing experiences can influence how we remember both our current experiences and unrelated subsequent experiences.