Mindware Ch 4: "Connectionism"

Theme of the reading
For the first 30 years of cognitive science (from the mid 1950s till the mid 1980s), there was only one computational framework that could satisfy the need in cognitive science to model reason-respecting behavior. That changed in the mid-1980s with the development of connectionism. Connectionism is "neurally inspired" computation. Rather than having a single processor that executes one instruction at a time, connectionist networks are composed of many "virtual" processors called units that are connected to each other by simple links, called connections, that pass activation or excitation between units. Rather than representing a concept by a discrete symbol, connectionist networks use "distributed" representations in which a representation of a concept is a pattern of activation across a group of units. Connectionism has many virtues as a computational framework. For example, some kinds of learning, generalization, and pattern completion are modeled very naturally in connectionist networks. It also has some important drawbacks. For example, it is often very difficult to determine how a connectionist network does what it does. And while "units" are vaguely like neurons, many key aspects of connectionist models do not fit well with what is known about how real brains operate.

Reading
Orienting questions and issues to keep in mind:
What is "graceful degradation"?
What methods are used to determine how connectionist networks do what they do?
What makes second-generation connectionism different from first generation connectionism?
Why does Clark call connectionism an unfinished revolution?