I have been reading the Liquid State Machine (LSM) literature, and I believe I have now grasped the main concepts behind it (embedding spike streams into high dimensionality, tuning the machine near criticality aids perfomance etc).
However, if one was to ask me "how many neurons do you need for a machine" I would not know. While the literature has been remarkably clear on the abstract concepts involving the operational principles behind LSM, I have not come across anyone explaining the trade offs of design.
Do more neurons help? What fraction of the neurons should be input neurons? How do I ensure critical-like dynamics for a specific system?
Is there literature on these questions?