this can be easily argued by looking at the demos of stochastic ANNs used for various perception tasks. for example: http://www.youtube.com/watch?v=0LTG64s6Xuc
patterns of characters images that emerge during simulations can be thought of as symbols
You think of a heap of stones as symbols too. The question is whether you can implement a symbolic system with it or not. The answer is no.
The paper discusses what you need for a symbolic system, and why it doesn't exist in the brain.