Parallel Distributed Processing and Other Computational Models: Strengths, Weaknesses, and Possibilities

Computational models have not found a place in either mainstream preclinical science or in the clinical science of cognitive function. In an effort to address this problem, I review the major computational models as they apply to brain function. These include parallel distributed processing (PDP) models, integrate and fire networks, brain in silico models, convolutional networks, statistical models, and others. I discuss PDP models in the greatest detail because they have most successfully accounted for a wide variety of behaviors in individuals who are healthy and those with brain damage. After discussing the strengths and weaknesses of all major models, I make proposals for further research. Due to the breadth of this review, a glossary of specialized terms is included in the Appendix, Supplemental Digital Content 1, http://links.lww.com/CBN/A155. The learning algorithms used in these models are problematic, including backpropagation in PDP models and unsupervised learning in others. Therefore, I review the extensive evidence that acetylcholine, delivered by the basal forebrain nuclei, both provides the crucial signal that Hebbian learning is to occur and enables that learning. Afferent input to these nuclei largely comes from the networks that represent subjective value, such as those in the orbitofrontal cortex. Of particular relevance to PDP models, a mechanism has recently been proposed by which backpropagation might occur in situ-triggered by acetylcholine-resulting in learning by virtue of the differential processing that occurs in dendrites and axons.