Shaping embodied neural networks for adaptive goal-directed behavior

Area of Science:

• Neuroscience
• Computational Biology
• Artificial Intelligence

Background:

• Learning and memory arise from synaptic modifications influenced by sensory feedback.
• The interplay between synaptic processes, neuronal population dynamics, and environmental interaction remains incompletely understood.

Purpose of the Study:

• To investigate how simulated neural networks can learn and adapt through a sensory-motor loop.
• To explore the role of synaptic plasticity in behavioral adaptation within a biologically inspired model.
• To develop a closed-loop hybrid system for studying network properties and behavioral adaptation.

Main Methods:

• Embodied a simulated neural network (inspired by cortical cultures) as an artificial animal (animat).
• Implemented a sensory-motor loop with structured stimuli, spatial activity metrics, and a spike-timing-dependent plasticity (STDP) training algorithm.
• Utilized adaptive stimulus selection based on animat behavior for training.

Main Results:

• The network learned associations between sensory inputs and motor outputs.
• The animat adapted to new sensory mappings to maintain goal-directed behavior (e.g., staying within a defined area).
• Successful learning depended on appropriate stimulus encoding and varied training stimuli.
• Networks exhibited flexibility in achieving multi-task goals and showed that different synaptic strengths could yield the same behavior.
• Biologically inspired networks tuned activity in behaviorally relevant ways, demonstrating innate information processing capabilities.

Conclusions:

• Leaky integrate-and-fire neural networks possess inherent information processing abilities.
• The closed-loop hybrid system is valuable for studying synaptic plasticity and behavioral adaptation.
• The adaptive training algorithm offers a foundation for future control systems in artificial and biological domains.