Learning by stimulation avoidance: A principle to control spiking neural networks dynamics.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• Current learning rules for neural networks lack general applicability.
• Biologically inspired models require robust learning principles for widespread use.

Purpose of the Study:

• To introduce and validate a new learning principle, "Learning by Stimulation Avoidance" (LSA), for biologically inspired neural networks.
• To demonstrate LSA's efficacy in reproducing biological learning results and explaining synaptic pruning.
• To show LSA's potential as a sensory-motor learning rule for embodied AI.

Main Methods:

• Simulating artificial neural networks with carefully timed external stimulation.
• Comparing simulation results with biological neuron learning data (Shahaf and Marom).
• Scaling network simulations from 3 to 100 neurons to examine underlying mechanisms.
• Applying LSA to a simulated robot for wall-avoidance learning.

Main Results:

• LSA effectively steers network dynamics towards desired states.
• The principle reproduces biological learning outcomes and explains synaptic pruning.
• LSA demonstrates higher explanatory power than existing hypotheses for biological neural network responses.
• Successful sensory-motor learning (wall avoidance) was achieved in a simulated robot using LSA.

Conclusions:

• Learning by Stimulation Avoidance offers a promising general learning rule for artificial neural networks.
• LSA provides a biologically plausible mechanism for learning and synaptic plasticity.
• This work presents a novel approach to Hebbian learning in spiking networks, independent of reward systems, for embodied applications.