Hebbian Descent: A Unified View on Log-Likelihood Learning