Evolving Plasticity for Autonomous Learning under Changing Environmental Conditions.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• Biological neural networks exhibit plasticity, enabling lifelong configuration changes.
• Hebbian learning offers a biologically plausible model for plasticity in artificial neural networks (ANNs) via local neuron interactions.
• The transition from local Hebbian rules to coherent global learning behavior remains poorly understood.

Purpose of the Study:

• To discover interpretable local Hebbian learning rules that facilitate autonomous global learning in ANNs.
• To investigate the emergence of global learning from local plasticity mechanisms.

Main Methods:

• Encoding learning rules using a discrete representation within a finite search space.
• Implementing synaptic changes based on local neuronal interactions.
• Utilizing genetic algorithms for rule optimization in online lifetime learning scenarios.
• Testing rules on foraging and prey-predator tasks.

Main Results:

• Evolved rules converged into distinct, interpretable types.
• The discovered rules demonstrated effective adaptation of ANNs during learning tasks.
• Performance of the evolved Hebbian rules was comparable to offline learning methods like hill climbing.

Conclusions:

• Interpretable local Hebbian learning rules can drive autonomous global learning in ANNs.
• Genetic algorithms are effective in discovering such rules for lifelong learning.
• This approach bridges the gap between local plasticity and global emergent behavior in artificial learning systems.