Adaptive Resonance Theory Mohamad DankarAUL hamra [email protected]Abstract— The reason behi nd ART model is tha t the identification of the objects and there recognition are the result of top-down interaction when observed with bottom-up sensory information. When compared with real object features as by the se ns es , ART mode l says that Top- down expe ctations are memory template form. It` s a descr ipt ion of a number of neur al network models in which we use supervis ed and unsup ervis ed learn ing meth ods, and where the addr ess proble ms are patter n rec ogniti on and prediction.. I. I NTRODUCTION Adapt ive resonance theory is devel oped by Carpe nter and Gro ssberg . It` s aim is for cluste ring binary vec tors wer e different models are found ART1 and ART2. ART1 accepts continuous-valued vectors .ART2 uses unsupervised learning metho d recog nition and predi ction inputs are in any orderweights are considered as code vectors for the cluster pattern. The difference between sensation and expectation should not exceed a set of threshold called the "vigilance p arameter". If not exceeded , sensed object is considered as a memberof the expected class. As mentioned before , ART system is an unsup ervis ed learn ing model of a compa rison fie ld and a rec ogn ition fie ld both compos ed of neu rons, a vigilance par ameter, and fin al ly a rese t module . The Vi gil ance parameter is highly influenceable on the system . For that , with high vigilance , highly detailed memories will be produced. while with lower vigi lance , the results will be more general for the memories .The comparison field takes a one-dimensional array of values input vector and transfers it to its best matc h in the recogni tion field. Recog nition fie ld neuro n , (1) ea ch gene rates a n egati ve sign al to ea ch of the other recogniti on fie ld neu rons and inh ibi ts the ir out put acco rdingl y. In this way the recognitio n field exhibits lateral inhib ition, allowin g each neuro n in it to represent a cate gory to which input vectors are classified. After classifiying the input vector, reset module is compared to the str eng th of the recognit ion mat ch to the vig ilance para meter .If vigila nce threshold is met, train ing star ts.If not, we see if the match le vel does not meet the vi gil ance para meter, than firing recog nition neuron will be inhibi ted until a new input vector is applied; the training will start only when search proce dure is done. Talking about the seach procedure,reset function will disable neurons recognition one by one until the vigilance parameteris satisfied by a recognition match. If not commit ted recogn itio n neu ron’s mat ch mee ts the vigilance threshold, then an uncommitted neuron is committed and adjusted towards matching the input vector. II.TRAINING We have two training methods : A. Slow B. Fast In the Slow learning method ,the training degree of the recog nition neuron`s weights towards input vector will be calcu lated with di ffere ntial eq uatio ns to c ontinuous values and so it is dependent on the length of time where the input vector is presented. With fast learning, we use algebraic equations to calculate de gr ee of we ight ad justments while us ing the bi na ry value s.Thou gh fast learning effec tivnes s and efficienc y are for a va riet y of ta sks, the slow lear ni ng me tho d is more biolo gical ly reason able and is used with con tinuou s-time networks. III . ART1 Stands for Adaptive Resonance Theory 1. An Unsupervised Clustering of binary input vectors.It is the simplest variety ofART networks, accepting only binary inputs. The architecture for the ART1 consists of F1 units , F2 units and a unit forres et that wil l implement use r to control simila rity degree placed on the same cluster.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Generally , the membership function of the fuzzy set A, ,
is prescribed as follows:
where X is the universal set. In this paper, seven
sets of fuzzy models are used--PB (positive big), PM (positive medium), PS (positive small), ZE (zero), NS (negative small), NM (negative medium), and NB (negative big)for those inputs andoutput of fuzzy controller. (5)
Figure VI
VIII Conclusion
We can not predict a proper vigilance value, with the help of
fuzzy controller, the ART1 network might be insensitive for
given initial vigilance values. Plus , a Modified Fuzzy ART
has advantages over the ART1 that are a minimum processing