Cause of failure with somatic time constants in evolutionary training

Ascertain why evolutionary training of the modified leaky integrate-and-fire networks failed when time constants were implemented at the soma rather than at synapses (somatic time constants), under the same simulation conditions used for semi-temporal logic tasks.

Background

The paper primarily employs a neuron model in which decay is represented via dendritic (synaptic) time constants, and demonstrates that co-adapting temporal parameters such as delays and time constants can solve diverse semi-temporal logic tasks. To probe model variants, the authors conducted additional simulations (not reported in detail) where the time constants were moved to the soma (somatic time constants).

They report an inability to evolve successful solutions under the same training regimen and acknowledge uncertainty about the underlying reason, offering a tentative hypothesis involving the need for contrastive time constants across inputs. This raises a specific unresolved question about the mechanistic cause of failure when adopting somatic rather than dendritic time constants in evolutionary optimization.