The Failure of Sub-Symbolic AI

Though connectionist systems have been modelled since before the Turing test was developed, they were not very popular until the late 80s.
In 1986, the PDP books were published and there was a surge in interest in connectionist systems.
Aside- I use neural nets to refer to models based on neurons, where connectionist systems are just inspired by neurons. A feedforward net is connectionist, not a neural net.-
The connectionist work has been remarkably successful.
There are a host of different connectionist algorithms,
they are used in industrial applications,
the understanding of them has grown rapidly.
The parallel nature of them also gives us hope of building really powerful parallel computers well beyond the power that a serial machine could have.
They also give us hope of grounding our symbols.
The problem is that most work is on learning simple functions.
Currently, there is no sound way of combining them to build larger systems.