Attractor-Driven Hierarchical Memory for Simultaneous Associative and Sequential Pattern Learning

TL;DR: Let’s give neural networks the brain’s ability to store both “facts” and “stories” by embedding attractor mechanisms into hierarchical memory—so they can do both associative memory and sequence recall at multiple abstraction levels.

Research Question: How can hierarchical memory networks leverage the coexistence of point and cyclic attractors to simultaneously support associative recall and sequential pattern recognition at multiple abstraction scales?

Hypothesis: Embedding competitive attractor mechanisms within hierarchical memory layers enables networks to flexibly switch between recalling static patterns and retrieving learned sequences, mimicking cognitive dual-memory systems.

Experiment Plan: - Setup: Implement hierarchical memory networks with attractor dynamics (drawing from Huo et al., 2024), allowing each layer to maintain either point or cyclic attractors.

• Data: Use tasks requiring mixed recall (e.g., image set recall and ordered sequence prediction).
• Measurement: Assess recall accuracy, sequence generation, and flexibility. Analyze the emergence and competition of attractors at different scales.
• Expected Outcome: Such networks outperform standard hierarchical memories in mixed-memory tasks and exhibit resilience to “memory interference.”

References:

• Huo, J., Yu, J., Wang, M., Yi, Z., Leng, J., & Liao, Y. (2024). Coexistence of Cyclic Sequential Pattern Recognition and Associative Memory in Neural Networks by Attractor Mechanisms. IEEE Transactions on Neural Networks and Learning Systems.
• Wang, L., Yang, M., Li, C., Shen, Y., & Xu, R. (2021). Abstractive Text Summarization with Hierarchical Multi-scale Abstraction Modeling and Dynamic Memory. SIGIR Conference.