Search Papers | Poster Sessions | All Posters
Poster B134 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Tracking representational formats of intrusive memories using deep neural networks
Rebekka Heinen1 (), Malte Kobelt1, Nikolai Axmacher1; 1Ruhr University Bochum, Germany
Intrusive memories of traumatic events differ in their content and their quality (i.e., representational format) from voluntary memories. We investigated these representational formats in 22 participants using a trauma film paradigm with a subsequent resting period to collect memory intrusions during functional magnetic resonance imaging (fMRI). We employed a convolutional neural network (DNN) re-trained to identify emotions, and large language model to quantify visual and semantic format. Using representational similarity analysis on DNN features we observed higher similarities between trauma than between neutral clips in both the visual and the semantic model, indicating generalization across content. However, on a neural level, encoding of trauma-analog clips revealed more pronounced visual formats. Our next steps will be to employ the semantic model and to analyze the resting period containing memory intrusions.
Keywords: memory deep neural networks large language models representational similarity analysis