Retrieved-context models propose that the context at encoding helps guide memory search for later retrieval. These models account for phenomena such as the temporal proximity effect, where events that are encoded close in time are also more likely to be retrieved together later on. However, the association between two temporally adjacent memories can be weakened by an event boundary, which helps segment our continuous life experience into discrete episodes. We are interested in how episodic richness of naturalistic events is affected by temporal distance and the presence of event boundaries.
This project was presented at Context and Episodic Memory Symposium and Lake Ontario Visionary Establishment in 2020. It is led by Bryan Hong, a PhD student from Memory & Perception Lab at University of Toronto. This repository contains implementations I work on under Bryan's supervision.
Forrest Gump is a modern classic, so many participants came into the lab with previous knowledge of the movie. This is a common issue dealing with naturalistic stimuli. To account for this, you can use survival analysis to pre-process the data. You will need to record the time when they expose to the stimulus to perform left censoring. If this information is not available (sometimes people don't remember when they first watched Forrest Gump), you will use left truncation to remove it from the dataset. A big enough sample size is the key!
We categorized the data into four memory types before proceeding to statistical analysis using a modified version of Autobiographical Interview. A detailed explanation of our approach can be found here on page 3. We have a VBA script that counts the number of highlights and a Python script that parse text with line breaks into individual .txt file if you are processing data in a word doc like us!
I created Arc specifically for this project as it nicely presents all the data points on a horizontal line (which is a good analogy of the timeline of the movie). An arc connects two correlated events, and its weight represents the strength of correlation. Here I use chi-square test to adjust the line weights, but you can replace it with any statistical tool that you want to use.
Big thanks to Bryan Hong for being a supportive mentor! Special award goes to Tianyu Lu and Ziyad Edher for the VBA script, and to Rui Liang, Michael Chen, Adam Huang, Zhao Lian, and Tianyu (thanks again!) for their support when I am exploring machine learning.