Symbolic processing in specially crafted networks
Modelling symbolic activity requires connectionist architectures that are "handcrafted to produce symbolic-like processing". Symbolic processes don’t emerge from connectionist networks, as Smolensky claims, they must be explicitly built-in.
In future, connectionism will have to move beyond simple feedforward networks to accommodate richer brain and spine architectures.
Connectionism, as part of a "team of concepts and tools", should ultimately model such phenomena as single-trial learning, attention, multispeed learning rates, and working memory.
Walter Schneider, 1988, p.51.