Multi-layer network utilizing rewarded spike time dependent plasticity to learn a foraging task.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Single-layer neural networks with reward-modulated spike-time dependent plasticity (STDP) can learn simple foraging tasks.
• Advanced learning capabilities require more complex network architectures and plasticity rules.

Purpose of the Study:

• To demonstrate advanced pattern discrimination and continuous learning in a multi-layer spiking neural network.
• To investigate the role of different STDP mechanisms and homeostatic regulations in complex task solving.

Main Methods:

• Utilized a multi-layer spiking neural network with both reward-modulated and non-reward-modulated STDP.
• Implemented homeostatic regulation mechanisms including heterosynaptic plasticity, gain control, and activity normalization.
• Introduced synaptic noise to facilitate trial-and-error learning.

Main Results:

• A hidden layer with non-rewarded STDP enabled neurons to classify input patterns based on specific input combinations.
• A subsequent layer with rewarded STDP allowed for discrimination between rewarding and punishing patterns without labeled data.
• Synaptic noise aided in identifying effective goal-oriented strategies for task completion.

Conclusions:

• Multi-layer spiking neural networks with diverse STDP rules and homeostatic mechanisms can achieve complex pattern classification and decision-making.
• The proposed network architecture and learning rules are sufficient for solving sophisticated foraging tasks.
• This work provides insights into the critical properties required for advanced learning in spiking neural networks.