From neurons to neural networks

Wednesday, April 2, 2008 - 11:15am - 12:15pm
Lind 409
Jeff Knisley (East Tennessee State University)
Artificial Neural Networks (ANN's) are machine-learning algorithms that are often used as classifiers in molecular and computational biology. Originally, ANN's were inspired by in vivo models of axonal and dendritic neuro-electric activity, especially the classical models of Hodkgin, Huxley, and others. Much of the successive development of ANN's, as well as the parallel development of other approaches such as Support Vector Machines, has been as a means of addressing issues such as overfitting and hard margins which arise in machine learning applications. To address these issues, ANN's have borrowed from a variety of sources in computer science, physics, and cognitive psychology, but not so much from the ever-improving neuronal models which provided their initial inspiration. We will revisit much of the historical and algorithmic development of ANN's, with the goal being that of suggesting the types of ANN's that might be inspired by more recent developments in dendritic electrotonic models.