A Two-Layer Self-Organizing Map with Vector Symbolic Architecture for Spatiotemporal Sequence Learning and...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spatiotemporal data presents challenges for current learning algorithms.
• Real-time prediction requires adaptive and flexible learning methods.
• Cognitive neuroscience inspires new approaches to pattern recognition.

Purpose of the Study:

• To propose a novel nature- and neuro-science-inspired algorithm for spatiotemporal learning and prediction.
• To enable fast and adaptive learning with dynamic data changes and concept drift.
• To improve the prediction of next occurrences in real-life spatiotemporal data.

Main Methods:

• Developed a spatiotemporal learning algorithm based on sequential recall and vector symbolic architecture (VSA).
• Decoupled learning of spatial and temporal patterns, using spatial patterns as an alphabet for temporal sequences.
• Employed unsupervised learning for pattern detection in unlabeled data streams.
• Utilized hyper dimensional (HD) computing-based associative memory for continuous prediction.

Main Results:

• The proposed ST-SOM algorithm demonstrated significant advantages over state-of-the-art methods in spatiotemporal unsupervised sequence learning.
• Achieved a 45% error reduction compared to the Hierarchical Temporal Memory (HTM) algorithm.
• Empirically evaluated using benchmark and time-series datasets.

Conclusions:

• The novel algorithm offers a flexible and adaptive approach for real-time spatiotemporal learning and prediction.
• Decoupling spatial and temporal patterns enhances adaptability to changing data and concept drift.
• The algorithm effectively addresses computational requirements for predicting sequential patterns.