Concept Learning, Recall, and Blending with Regulated Activation Networks
Authors
Abstract
Herein we present the cognitive model Regulated Acti-vation Networks (RANs), which aims at unifying the three
perspectives (symbolic, connectionist, and geometric feature-
space) of conceptual representations. It learns new concepts
from input data, dynamically builds a hierarchy of abstract
concepts, and learns the associations among them, both be-
tween different levels, and within the same level of the hier-
archy. Its recall mechanism, the geometric backpropagation
algorithm, allows the understanding of the meaning of higher
level concepts in terms of input level features. The regulation
mechanism we also introduce has a de-noising effect over the
results obtained from the recall mechanism.