Training a spiking neuronal network model of visual-motor cortex to play a virtual racket-ball game using...

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Robotics

Background:

• Spiking neuronal networks (SNNs) show promise for complex behaviors but often struggle in dynamic environments.
• Existing SNN models lack analysis of circuit architectures and learning mechanisms for optimal performance.
• Testing SNNs in dynamic environments requires predictive capabilities and adaptive behaviors.

Purpose of the Study:

• To develop and evaluate visual/motor SNN models for dynamic environment tasks.
• To investigate the impact of different circuit architectures (feed-forward, recurrent, feedback) on learning and performance.
• To create and validate a novel, biologically-inspired learning rule to improve SNN efficiency.

Main Methods:

• Trained SNN models to play a virtual racket-ball game using reinforcement learning algorithms.
• Systematically analyzed the contribution of various circuit motifs to learning and performance.
• Developed and implemented a new biologically-inspired learning rule.
• Modeled visual processing, association areas, and motor control pathways within the SNN.

Main Results:

• The developed SNN models successfully learned to play the racket-ball game in a dynamic environment.
• A novel biologically-inspired learning rule significantly improved performance and reduced training time.
• Analysis revealed that specific circuit architectures (e.g., recurrent, feedback) are crucial for effective learning.
• The model demonstrated how different neural circuits contribute to sensory-motor tasks.

Conclusions:

• Biologically-plausible learning rules are effective for training SNNs in dynamic environments.
• Understanding circuit architectures and learning rules is key to optimizing SNN performance.
• Biological systems exhibit resilience and redundancy through complementary learning mechanisms and neural circuits.