Self-organizing maps based on limit cycle attractors.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Conventional self-organizing maps (SOMs) often use static representations, which are inconsistent with the brain's dynamic, rhythmic activity.
• Limited exploration of SOMs in large-scale neurocognitive architectures, partly due to representational limitations.

Purpose of the Study:

• To develop and investigate an alternative encoding scheme for SOMs using sparsely-coded limit cycles.
• To explore the reliability, robustness, and associative memory potential of limit cycle-based SOMs.

Main Methods:

• Developed a novel SOM model utilizing limit cycles for input pattern representation.
• Analyzed the conditions for reliable limit cycle emergence and dominance in SOM dynamics.
• Investigated a two-SOM architecture to assess hetero-associative memory capabilities.

Main Results:

• Learned limit cycle representations were reliably formed and dominated SOM dynamics.
• These limit cycles demonstrated uniqueness for different inputs, robustness to perturbations, and timing insensitivity.
• A two-SOM system successfully evoked corresponding limit cycles, demonstrating hetero-associative memory potential.

Conclusions:

• Limit cycle-based SOMs offer a viable alternative to static representations, aligning better with neural dynamics.
• This approach shows promise for developing more sophisticated neurocognitive architectures and associative memory systems.
• Further research into limit cycle dynamics in SOMs is warranted.