Stimulus-to-stimulus learning in RNNs with cortical inductive biases

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Neuroscience

Background:

• Animals learn through conditioning, predicting external events via stimulus substitution.
• Neuronal responses to conditioned stimuli (CS) can become similar to unconditioned stimuli (US).

Purpose of the Study:

• To propose a recurrent neural network model for stimulus substitution.
• To investigate the role of cortical inductive biases in associative learning.

Main Methods:

• Developed a recurrent neural network model incorporating representational and architectural inductive biases.
• Modeled two-compartment pyramidal neurons as fundamental units for associative learning.
• Implemented a biologically plausible, locally-gated learning rule for stimulus substitution.

Main Results:

• The model successfully generates diverse conditioning phenomena.
• It learns numerous associations with training comparable to animal experiments, avoiding parameter fine-tuning.
• Contrasted with Hebbian rules, which struggle with mixed selectivity and require task-specific tuning.

Conclusions:

• Multi-compartment neuronal processing is crucial for cortical associative learning.
• This framework provides a biologically plausible mechanism for stimulus substitution and may confer evolutionary advantages.