Conditioning in hetero-associative neural networks trained with three-factor predictive plasticity

Animals learn from experience the value of environmental stimuli and adjust their behavior accordingly. At the core of animal conditioning lies the capacity to associate a neural activity pattern induced by an unconditioned stimulus (US) with the pattern arising in response to a conditioned stimulus (CS). Reward-modulated associative synaptic plasticity has been successful in explaining conditioning when the neural representations of behavioral stimuli are unmixed. However, this assumption is inconsistent with the fact that neurons --- particularly in high-level, cognitive areas --- display mixed selectivity.

Inspired by experimental findings on the associative power of single cortical pyramidal neurons, we propose a computational model that achieves generic pattern-to-pattern mappings at the population level. Our model incorporates a local learning rule operating in compartmentalized neurons, which mirrors the capacity of cortical pyramidal neurons to implement predictive learning through coincidence detection. These properties confer crucial advantages over three-factor Hebbian plasticity in point neurons and allow our model to account for a wide gamut of conditioning phenomena, including trace and delay conditioning, extinction, reported neuromodulator dynamics, the S-shaped acquisition curve characteristic of animal conditioning and the drop in conditioning effectiveness with larger delays between the CS and the US. Allowing for the simultaneous presentation and competition of CSs naturally gives rise to phenomena like blocking, overshadowing, saliency effects and overexpectation. Finally, examining the impact of CS contingency on conditioning we find that the model offers a reductionist mechanism for causal inference by resolving the post hoc fallacy, which states that when event Y occurs after event X, then X is considered its cause.

The model makes testable predictions about the evolution of neural mixed representations during conditioning experiments that might prove useful to understand existing datasets and guide future experiments. In phychological terms, it corresponds to the stimulus substitution component of classical conditioning.