Fast heterosynaptic learning in a robot food retrieval task inspired by the limbic system

Area of Science:

• Computational Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Hebbian learning, a prominent correlation-based learning paradigm, strengthens synapses with coinciding pre- and postsynaptic activity.
• Classical Hebbian learning exhibits instability due to an autocorrelation term, leading to exponential weight growth.
• Existing solutions often involve compensating for this autocorrelation term.

Purpose of the Study:

• To introduce a novel heterosynaptic learning rule that overcomes the instability inherent in classical Hebbian learning.
• To present a learning rule free from autocorrelation terms, ensuring greater synaptic weight stability.
• To demonstrate the efficacy of the proposed learning rule in a biologically inspired model.

Main Methods:

• Development of a heterosynaptic learning rule designed to avoid autocorrelation.
• Implementation and testing of the novel learning rule in a computational model.
• Utilizing a model inspired by the limbic system for evaluating learning performance in a food retrieval task.

Main Results:

• The proposed heterosynaptic learning rule demonstrates significantly greater stability compared to classical Hebbian learning.
• The absence of an autocorrelation term prevents the exponential weight growth characteristic of Hebbian learning.
• The learning rule proved effective in the limbic system-inspired food retrieval model.

Conclusions:

• The novel heterosynaptic learning rule offers a more stable and robust alternative to classical Hebbian learning.
• This approach eliminates the instability issue by removing the autocorrelation term.
• The demonstrated performance in a biologically relevant model suggests potential applications in understanding neural plasticity and developing advanced AI systems.