Biological neuronal systems are sloppy and imprecise but also extremely flexible and adaptable in navigating the environment by closing the loop between sensors and actuators and by exploiting ongoing error signals for online self-tuning.

 

Could the lack of precision and the mismatch problem of analog neuromorphic approaches be an artifact of having focused too much on comparisons with classical machine learning methods in disembodied benchmarks? Could they be overcome altogether, and possibly exploited, by bringing neuromorphic intelligence at the service of control engineering of highly uncertain and/or extremely low power physical plants?