A Proposal and Implementation of a Neural Network Based Hierarchical Temporal Memory to Realize Cognitive Functions
No Thumbnail Available
Date
2008-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
A Proposal and Implementation of a Neural Network Based Hierarchical Temporal Memory to Realize Cognitive Functions
Abstract
Hierarchical Temporal Memory (HTM) is a recent innovation in cognition science. Developed
in 2005 by Numenta Inc., an artificial intelligence research firm in the US, HTMs attempt to
capture the way the human brain learns and infers its environment. One of the most notable
characteristics of this model is the consideration of the hierarchical organization of objects in
the world. Data in the world is made up of elementary features that aggregate in successive
layers to form perceivable objects. This data can be visual, auditory or from other abstract
spaces such as stock markets and scientific studies. The amount of raw data that the brain is
exposed to throughout its lifetime is beyond imagination. However the brain is known to use a
very noble and systematic approach to handle the perception, storage, and inference of this
data. Several studies in neuroscience and psychology indicate that the brain makes use of the
hierarchies that features in the world exhibit in their organization to form objects. Hence, for
instance, ‘corners’ and ‘lines’ can aggregate to form a ‘table’ object in the visual world. These
elementary features, however, can use a different aggregation to form a ‘chair’ object. The
same is true for data in other types of worlds such as audio. HTMs directly apply a similar
handling of world data for their cognition. Furthermore, the structure of HTMs, made up of
data processing nodes arranged in a hierarchical tree, mimic the physical arrangement of
cortical layers in the brain.
The various data analysis algorithms in the nodes of HTMs were, however, found by the
researcher to be limiting in several aspects, one of which is handling of unforeseen (untrained)
data. For this purpose neural network data structures were used to replace some of the
operations in these nodes for their ability in approximating untrained data to nearest matches.
This research work discusses the proposed model, implementation constraints, the operational
characteristics, and performance enhancements observed in this modified model with a
selected test application. A substantial improvement in cognitive functionality has been
observed with the newly proposed model.
Description
Keywords
Hierarchical, Temporal