> See Hebbian learning

The one that is not used, because it's inherently unstable?

Learning using locally accessible information is an interesting approach, but it needs to be more complex than "fire together, wire together". And then you might have propagation of information that allows to approximate gradients locally.