Do connectionist networks accurately reflect neural computations?
Connectionist networks used in cognitive science research and engineering are often simplified versions of biological neural networks.
- Nodes: Do individual nodes in connectionist networks correspond to individual neurons, or groups of neurons?
- Coding: What does activation activity correspond to? Firing rate of neurons?
- Weight changes: In connectionist networks, connections can change from excitatory to inhibitory depending on the direction of weight change. But neural synapses are exclusively of one type only, i.e. either excitatory or inhibitory but not both.
- Learning: Typical learning algorithms such as BackPropagation are not biologically realistic.