Two-shot learning of multiple strange attractors.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Complex Systems

Background:

• The brain's ability to integrate short- and long-term memory is crucial for information processing.
• Recent advancements in multifunctional and parameter-aware learning inspire new computational models.
• Processing and recalling multiple dynamic systems, like strange attractors, presents a significant computational challenge.

Purpose of the Study:

• To develop and evaluate a novel machine learning system that mimics the brain's memory integration capabilities.
• To process, store, and recall multiple different strange attractors using a combined computational approach.
• To improve the accuracy and stability of memory recall in complex dynamic systems.

Main Methods:

• A hybrid machine learning system integrating a next-generation reservoir computer (NGRC) with extremely randomized trees (ERT).
• Training the NGRC+ERT system using a two-shot learning approach for efficient feature selection and reduced hyperparameter tuning.
• Utilizing an exponential filtering scheme for accurate reconstruction of short- and long-term dynamics.

Main Results:

• The NGRC+ERT system accurately reconstructed the dynamics of Lorenz and Halvorsen chaotic attractors.
• The system successfully processed and recalled 16 different attractors, demonstrating stability through feature space separation.
• Identified a link between short-term memory processing defects and long-term memory recall failures (confabulation).

Conclusions:

• The combined NGRC+ERT system offers a powerful tool for processing and recalling complex dynamic systems.
• The two-shot learning approach enhances performance and reduces computational overhead.
• Understanding memory processing failures in artificial systems can provide insights into biological memory mechanisms.