Adaptive structure evolution and biologically plausible synaptic plasticity for recurrent spiking neural networks.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Human-like intelligence relies on brain architecture and multi-scale learning.
• Liquid State Machines (LSMs) are brain-inspired models suitable for studying intelligence.
• Current LSM research often overlooks brain's evolutionary and learning mechanisms.

Purpose of the Study:

• To present a novel LSM learning model integrating adaptive structural evolution and multi-scale biological learning rules.
• To address limitations in current LSM research by incorporating brain's evolutionary and learning mechanisms.
• To enhance LSMs for complex decision-making tasks through biologically inspired design.

Main Methods:

• Developed an adaptive evolvable LSM model for optimizing liquid layer architecture.
• Proposed a dopamine-modulated Bienenstock-Cooper-Munros (DA-BCM) method for brain-inspired learning.
• Incorporated global dopamine regulation and local trace-based synaptic plasticity.

Main Results:

• Structural evolution of the liquid layer improved LSM decision-making.
• DA-BCM regulation enhanced LSM adaptability, including rule reversal.
• The integrated model demonstrated improved performance on decision-making tasks.

Conclusions:

• Adaptive structural evolution and DA-BCM learning rules significantly improve LSM performance.
• The proposed model offers a more biologically plausible approach to artificial intelligence.
• This work highlights the potential of evolutionary and neuroplasticity principles in designing advanced AI systems.