A Neural Model for Self Organizing Feature Detectors and Classifiers in a Network Hierarchy
Williamson, James R.
MetadataShow full item record
Many models of early cortical processing have shown how local learning rules can produce efficient, sparse-distributed codes in which nodes have responses that are statistically independent and low probability. However, it is not known how to develop a useful hierarchical representation, containing sparse-distributed codes at each level of the hierarchy, that incorporates predictive feedback from the environment. We take a step in that direction by proposing a biologically plausible neural network model that develops receptive fields, and learns to make class predictions, with or without the help of environmental feedback. The model is a new type of predictive adaptive resonance theory network called Receptive Field ARTMAP, or RAM. RAM self organizes internal category nodes that are tuned to activity distributions in topographic input maps. Each receptive field is composed of multiple weight fields that are adapted via local, on-line learning, to form smooth receptive ftelds that reflect; the statistics of the activity distributions in the input maps. When RAM generates incorrect predictions, its vigilance is raised, amplifying subtractive inhibition and sharpening receptive fields until the error is corrected. Evaluation on several classification benchmarks shows that RAM outperforms a related (but neurally implausible) model called Gaussian ARTMAP, as well as several standard neural network and statistical classifters. A topographic version of RAM is proposed, which is capable of self organizing hierarchical representations. Topographic RAM is a model for receptive field development at any level of the cortical hierarchy, and provides explanations for a variety of perceptual learning data.