|
MARF's Pattern Recognition Pipeline
|
Pattern recognition is a widely used biologically inspired technique in the modern computer science. Algorithms for image
and voice recognition have been derived from the human brain, which uses pattern recognition to recognize shapes, images,
voices, sounds,
avors, odors, etc. In this research, we applied the principles of AC(autonomic computing) to solve speciic
problems in pattern-recognition systems, such as availability, security, performance, and coniguration management. We
tackled these issues by using ASSL to introduce self-management in the system behavior.
As a proof of concept case study, we develop self-managing autonomic properties for DMARF (Distributed Modular
Audio Recognition Framework). DMARF is based on the classical MARF whose pipeline stages were made into distributed nodes.
The Modular Audio Recognition Framework (MARF) is an open-source research platform and a collection of pattern-recognition,
signal-processing, and natural language-processing (NLP) algorithms written in Java and arranged into a modular and extensible
framework facilitating addition of new algorithms for use and experiments by scientists. MARF has a number of algorithms implemented
for various pattern recognition and some signal processing tasks. Conceptually, it consists of pipeline stages that communicate
with each other to get the data they need in a chained manner. Four core stages groupe similar kinds of algorithms:
(1) sample loading, (2) preprocessing, (3) feature extraction, and (4) training/classification.
|
We specify with ASSL a number of autonomic properties for DMARF, such as self-healing, self-optimization,
and self-protection. The implementation of those properties is generated automatically by the ASSL framework in the
form of a special wrapper Java code that provides an autonomic layer implementing the DMARF's autonomic properties. In addition,
the latter are formally validated with the ASSL's mechanisms for consistency and model checking.
|