The study investigates the capacity of working memory to retain past visual information under continuous visual stimuli using fMRI signals. The authors employ ridge regression analysis and trial-wise representational similarity analysis (RSA) to assess the correlation between fMRI signals and visual stimuli from different past time points. They find that the correlation between fMRI signals and past semantic information gradually decreases over time, retaining at most 3-4 items, which aligns with the characteristics of working memory.
Based on these findings, the authors propose the "Memory Disentangling" task, which aims to extract past visual stimuli information from brain activity and separate it from ongoing brain activity to mitigate the effects of proactive interference. They introduce a disentangled contrastive learning method inspired by the phenomenon of proactive interference to accomplish this task. The method separates the information between adjacent fMRI signals into current and past components and decodes them into image descriptions.
Experimental results demonstrate that the disentangled contrastive learning method effectively disentangles the information within fMRI signals, improving the decoding of current visual information compared to a straightforward approach. However, the method's effectiveness in extracting past memory information is suboptimal, indicating the need for further exploration to optimize the model's ability to accurately capture past memory representations.
Іншою мовою
із вихідного контенту
arxiv.org
Ключові висновки, отримані з
by Runze Xia, C... о arxiv.org 10-01-2024
https://arxiv.org/pdf/2409.20428.pdfГлибші Запити