The Role Of Pattern Separation And Working Memory Abilities In Different Community Structures Of Non-spatial Graph Learning - Info and Reading Options

The Internet Archive:

The current project aims to investigate the association between an individual’s pattern separation abilities with their “Graph maze” task performance. Pattern separation is the ability to keep similar or overlapping memories separate in your mind, such as where you parked your car today vs where you parked your car yesterday. Participants, aged 18-35, who passed quality control and were included in the initial preliminary analysis of the Graph Maze Task will be reinvited to participate in a follow up study that will include the Mnemonic Similarity Task (MST), which has been shown to be an effective means of characterizing an individual’s pattern separation abilities. Our initial experiments found that there were differences in graph learning performance based on the structure of the graph. There were also large individual differences in performance overall (Kapogianis, in preparation). Given that each graph in our initial study has a unique structure and node organizational pattern, each stimuli object group is unique. However, there is a possibility that there are commonalities between object groups since these groups presented are the adjacent nodes to the currently ‘occupied’ node. For example, if node A is adjacent to both nodes B & C, then node A will be presented when participants occupy either node B or C. These commonalities in node objects across presentations increase with more highly interconnected communities. This gives rise to more opportunity for false recall, which can lead to an incorrect representation of state space. Pattern separation could be important in this situation to help keep the common nodes and similar adjacencies distinct from each other. Graph-like representations of more abstract associations have been recently observed throughout the brain. Representations of state space in medial prefrontal cortex (mPFC) represented nodes within a community – a tightly interconnected cluster of nodes – more similarly within communities than between communities1. Entorhinal cortex has also been shown to code for elements of a graph, representing relative edge distances between nodes in an abstract graph2, while hippocampus was shown to code for semantic similarity between nodes for the same information3. Additionally, the orbitofrontal cortex was shown to code for state space of unobservable task-relevant information4, which could be important for goal-directed planning. Prior work indicates that individuals can extract underlying graph structure in more abstract relationships through experience4-7, such as learning which nodes are closely associated. This ability to extract and connect abstract information through experience can most clearly be observed in recent work which presented pairs of items to participants during the Graph Walk Task5. These item pairs were actually disjointed edge-pairs – two nodes with a common edge – from a larger graph which participants were naïve to, presented in a random order. After repeated presentations of these edge-pairs, participants were able to make accurate relative edge-distance estimations, as well as reconstruct the graph itself, even though they were naïve to the graph's underlying structure at the start of the experiment and they never saw the whole graph. These findings show that individuals learn and connect associations from disordered experiences without prompting or knowledge of the underlying structure during latent learning. Importantly, it was found that performance during the Graph Walk Task was associated with an individual’s pattern separation ability8, and that those with lesser pattern separation abilities could overcome performance deficits when stimuli pairs were presented in blocks to be non-overlapping. For example, if node A is adjacent to nodes B & C, the original stimulus presentations would include A-B and A-C pairs. High pattern separators would have less difficulty distinguishing between A-B and A-C pairs during the same block, while low pattern separators would have more difficulty. But by only presenting pair A-B during study block 1 and pair A-C during block 2, these low pattern separators were able to overcome performance deficits. Our preliminary data for the Graph Maze task suggests that elements of a graph’s underlying structure such as size, edge-to-node ratio, and community structure all impact task performance. However, we theorize that another contributing factor may be an individual’s ability to pattern separate between similar stimuli and whether they are focused on local vs global structure of the underlying graph. Incorporating information into its global context may make regular organizational patterns more apparent and reduce reliance upon local pattern separation by using predictable patterns. This would be similar to navigating a city with a grid layout. However, learning based on local structure would require more robust representations of local connections, and be more reliant on pattern separation abilities in situations with low discriminability at the node level. (see attached image for better table view) Random Grid-like Comm-X-bridge Comm-w/bridge Regularity (global) ↓ ↑↑ ↑ ↑ Discrim. (local) ↑ ↑↑ ↓↓ ↓ Table 1: Regularity (global patterns) and discriminability (local patterns) of graphs used in Community Structure experiment. Graphs with higher regularity have a more global organizational pattern to connections. Graphs with high discriminability have less overlap of object groups meaning each group is relatively unique. The Graph Maze Task has a single study block for each graph where participants are exposed to the underlying structure through self-determined, temporally linked presentations of stimuli which represent the connections of their prior choice. Given this setup, we propose alternative hypotheses for each experiment. The results of the present MST study will adjudicate between them. Parameter search experiment: All graphs in this experiment were randomly generated, so there is no designed community structure. Thus, the focus here is on node number and ENR. a. High pattern separators will have greater task accuracy on high ENR graphs than low pattern separators. Low pattern separators will have greater task accuracy on low ENR graphs than high pattern separators (similar to Noh et al., 2023). b. Alternatively, high pattern separators may have greater task accuracy overall. c. Alternatively, we may see no difference across pattern separators on small graphs and see differences in accuracy across high and low pattern separators on larger graphs similar to hypothesis a. d. Additionally, we may see that working memory may play a larger role as ENR rises as the average stimulus group size grows. In this case, people with greater working memory ability would perform better with greater ENR. Community structure experiment: Here we follow the ideas laid out in Table 1 regarding the global regularity and local discriminability. e. If pattern separation is more relevant to regularity in global structure than to local patterns, we anticipate that task accuracy will be correlated with MST score in order of graph regularity (from lowest to highest performance: random, community w/ bridge, community no bridge, grid-like) f. If pattern separation is more relevant to discriminability in local patterns than to global structure, we anticipate that task accuracy will be correlated with MST score in order of graph discriminability (from lowest to highest performance: community no bridge, community w/ bridge, random, grid-like) g. Alternatively, trial edge distance (how far apart the start and target locations are from each other in terms of number of nodes) may be a significant factor for pattern separation and may need to be taken into consideration during analysis. Trial distance may determine whether start and target nodes are within or across communities, which could affect the degree to which global patterns are needed on any given trial. h. Additionally, we may see that working memory may play a greater role in graphs with greater community structure because participants are forced to plan more. Thus, we would expect the greatest relationship between working memory performance in task performance in the community with bridges and communities without bridges graphs. To address these hypotheses we will invite participants who completed the Graph Maze Task, for both our parameter search and community structure experiments, through Prolific to complete the MST, Spatial N-back, and Verbal N-back, also on Prolific. We will run a linear mixed model of MST Lure Discrimination Index scores, N-back Scores, node number, and edge-to-node ratio with Graph Maze task performance to evaluate the relationship between pattern separation ability, graph structure, and task performance, as well as the contributions of working memory. Participants will be invited to participate through Prolific, and will receive 3 weekly reminders following their initial invitation. The study will remain open for invited participants until 1 week after the final reminder or our goal of 80% of invitees complete the task. The parameter search Graph maze task had 208 participants, so we will target 166 participants. The community structure graph maze task had 52 participants, so we will target 42 participants. Graph Maze Task Description: Participants will complete four blocks of an N-alternative forced choice “graph maze” task, each with a unique non-spatial graph. Each block will consist of an exploration phase and a test phase. Exploration: Before each exploration phase participants will be informed that they will be presented with groups of up to three objects that they must memorize for a subsequent memory test. The language of the instructions will be carefully chosen to exclude spatial and navigation language to prevent priming of a spatial/ navigation reference. They will proceed through these object groups – “navigate” the graph whose underlying graph structure they are naïve to – at a self-guided pace until their exploration time ends. To proceed to the next object group, an item must be selected. To encourage novel object selection, participants will be informed that making different selections will result in different groups being presented. The objects displayed on the screen represent the adjacent (connected) Nodes to the currently occupied node. By selecting one object, participants will transition to a new node and the object group presented will be updated. This process repeats itself until the exploration time for that graph ends. Test: During the test phase, participants are asked to select the target item when it appears, by proceeding through the object groups the same way as in the exploration phase. The target item and the previously selected object will be displayed on the screen during the test trials to reduce working memory load. Participants complete a total of 20 trials per graph, each with a 20s trial time. Mnemonic Similarity Task Description: The task contains two phases: encoding and test phases. The encoding phase consists of a cover task which results in incidental encoding where participants judge whether stimuli are “indoor” or “outdoor” items. Immediately following the encoding phase, participants will perform a recognition task during the test phase, where they will judge stimuli as either new, similar, or old. One-third of the images in the test phase are exact repetitions of images presented in the study phase (targets); one-third of the images are new images not previously seen (foils); and one-third of the images are perceptually similar to those seen during the study phase, but not identical (lures). Lure Discrimination Index (LDI) scores are calculated by calculating the difference between the probability of giving a “Similar” response to the lure items minus the probability of giving a “Similar” response to the foils to account for any bias the participant may have in using the “Similar” response overall. N-back Task Description: The N-back is a test of working memory. It asks participants to continuously update task relevant information and relate it to the current stimulus. The task contains three distinct practice sessions of increasing difficulty and one test phase. Participants are shown a series of stimuli, and asked to respond when the current stimulus matches the stimulus N places back. For example in the 1-back block if the participant saw the sequence of letters “T-K-K”, they would not respond at “T’ or first “K” because those do not match stimuli 1 place back. However, they would respond to the second “K” shown, because it matches the stimulus 1 place back. There are 1-back, 2-back, and 3-back practice blocks and participants must have greater than 80% accuracy on the 1-back and 2-back blocks to proceed. Responses are categorized as a hit, correct rejection, false alarm, or miss. The N-back scores are calculated by the proportion of correct responses (hit and correct rejection) to the total number of stimuli presented. N-back scores give us a quantitative measure of someone’s working memory capacity in a task environment which is continually updating. Both verbal (using letters) and spatial (using locations on a 4x4 grid) versions are used. Analysis Plan: Parameter search: We will conduct a Linear Mixed Model (LMM) between LDI scores from the MST, Spatial and Verbal N-back scores, and the proportion of correct trials and path efficiency (excess steps) during the test phase of the Graph Maze Task. In addition to finding the raw correlation values, we will compare the correlation effect sizes to determine whether there are statistically significantly greater correlations in some conditions of the Graph Maze Task (node number and ENR of graphs). Community Structure: We will conduct a Linear Mixed Model (LMM) between LDI scores from the MST, Spatial and Verbal N-back scores, and the proportion of correct trials and path efficiency (excess steps) during the test phase of the Graph Maze Task. In addition to finding the raw correlation values, we will compare the correlation effect sizes to determine whether there are statistically significantly greater correlations in some conditions of the Graph Maze Task (grid-like, random, community with bridges, community without bridges) than in others. Schapiro, A. C., Rogers, T. T., Cordova, N. I., Turk-Browne, N. B. & Botvinick, M. M. Neural representations of events arise from temporal community structure. Nature Neuroscience 2013 16:4 16, 486–492 (2013). Garvert, M. M., Dolan, R. J. & Behrens, T. E. A map of abstract relational knowledge in the human hippocampal-entorhinal cortex. Elife 6, (2017). Zheng, X. Y. et al. Parallel cognitive maps for short-term statistical and long-term semantic relationships in 1 the hippocampal formation Correspondence to. bioRxiv (2023) doi:10.1101/2022.08.29.505742. Schuck, N. W., Cai, M. B., Wilson, R. C. & Niv, Y. Human Orbitofrontal Cortex Represents a Cognitive Map of State Space. Neuron 91, 1402–1412 (2016). Rmus, M., Ritz, H., Hunter, L. E., Bornstein, A. M. & Shenhav, A. Humans can navigate complex graph structures acquired during latent learning. Cognition 225, (2022). Schapiro, A. C., Rogers, T. T., Cordova, N. I., Turk-Browne, N. B. & Botvinick, M. M. Neural representations of events arise from temporal community structure. Nature Neuroscience 2013 16:4 16, 486–492 (2013). Kahn, A. E., Karuza, E. A., Vettel, J. M. & Bassett, D. S. Network constraints on learnability of probabilistic motor sequences. Nature Human Behavior 936–937 (2018) doi:10.1038/s41562-018-0463-8. Noh, Sharon M., Keiland W. Cooper, Craig Stark, and Aaron Bornstein. Multi-step Inference Can Be Improved Across the Lifespan with Individualized Memory Interventions. PsyArXiv (2023) doi:10.31234/osf.io/3mhj6.

Search for “The Role Of Pattern Separation And Working Memory Abilities In Different Community Structures Of Non-spatial Graph Learning” downloads:

Visit our Downloads Search page to see if downloads are available.

Find “The Role Of Pattern Separation And Working Memory Abilities In Different Community Structures Of Non-spatial Graph Learning” in Libraries Near You:

Read or borrow “The Role Of Pattern Separation And Working Memory Abilities In Different Community Structures Of Non-spatial Graph Learning” from your local library.

• The WorldCat Libraries Catalog: Find a copy of “The Role Of Pattern Separation And Working Memory Abilities In Different Community Structures Of Non-spatial Graph Learning” at a library near you.