Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Attention rules Episodic Memory

Zahra Fayyaz¹, Sen Cheng², Laurenz Wiskott¹; ¹Ruhr-Universität Bochum, ²Ruhr-Universtät Bochum

Presenter: Zahra Fayyaz

Attention plays a crucial role in memory and learning by prioritizing relevant information and filtering out redundant input. This study explores how attention, guided by semantic memory, enhances memory encoding and retrieval. We present a neural network model to simulate generative episodic memory, comprising a VQ-VAE encoder, an attention module, and a transformer-based semantic decoder. Three attention strategies (random, selective, and additive) were evaluated. Random attention, lacking prioritization, led to lowest memory accuracy. Selective attention, informed by semantic prediction, improved performance by focusing on novel, informative inputs. Additive attention, inspired by biological saccades, offered the highest performance through iterative, predictive refinement of input encoding, albeit at a higher computational cost. Furthermore, experiments on both MNIST and ImageNet datasets demonstrate that semantically-guided attention leads to more structured and less prototypical memory traces. These findings underscore the dynamic interplay between attention and memory, suggesting that attentional mechanisms shaped by prior knowledge significantly optimize learning and memory.

Topic Area: Object Recognition & Visual Attention

Extended Abstract: Full Text PDF