Sejnowski and Rosenberg observe that the context (7 wide) of a given
letter is enough to determine ``most'' of its pronunciation.
They code a character (26 letters, space and punctuation) as 29 units, and
an input layer
as 
units.
One hidden layer of 80 units connects to 26 output units that
determine a phoneme in distributed fashion.
The network has full connectivity between layers, and each unit has a threshold, so there are
degrees of freedom in this network.
On a limited training set (a recording of 1024 spoken words) the system learned surprisingly quickly, and could generalise to unseen text with good performance.
The network was damage resistant to the extent that reasonable performance could be maintained when the information content of each weight was reduced to 4 bits.
In a similar fashion to the preceding example, the hidden units adopted (distributed) behaviour that could be tied to the input/output patterns - in this case, activity corresponded to letter/sound pairs.