Precisely, connectionism in modern AI boils down to the idea that learning should be expressed in terms of DAGs that are composed of simpler units. It’s quite likely that the units that are currently used are too abstract, but this doesn’t necessarily mean the paradigm itself is flawed.
I don't think connectionism is restricted to acyclic graphs. Or even graphs in general. But you're right that the connectionism as an approach is more abstract than just simulating neuron behavior.
