Dynamic Network Plasticity and Sample Efficiency in Biological Neural Cultures: A Comparative Study with Deep...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• Investigating in vitro neural systems provides insights into complex network dynamics.
• Live neural cultures integrated with technology offer novel research platforms.
• Understanding neural plasticity is key to deciphering learning mechanisms.

Purpose of the Study:

• To analyze network dynamics in live neural cultures during gameplay.
• To compare the learning efficiency of biological neural systems with deep reinforcement learning (RL) algorithms.
• To introduce a framework for comparing biological and artificial neural network performance.

Main Methods:

• Utilized DishBrain, integrating live neural cultures with multi-electrode arrays in closed-loop game environments.
• Analyzed neural activity by embedding spiking data into lower-dimensional spaces.
• Compared performance of neural cultures and RL algorithms (DQN, A2C, PPO) in a Pong simulation.

Main Results:

• Distinguished between spontaneous and gameplay-driven neural activity patterns.
• Observed dynamic changes in neural connectivity, indicating sample-efficient plasticity.
• Biological neural cultures demonstrated superior performance over deep RL algorithms in limited sample conditions.

Conclusions:

• In vitro neural systems exhibit high sample efficiency in learning and adaptation.
• Biological neural networks offer a valuable benchmark for artificial intelligence development.
• DishBrain facilitates real-time monitoring and manipulation of neural network dynamics.