An Analysis on Pattern Recognition Using Machine Learning

I. INTRODUCTION

A pattern "it is a substance that could be given a name". For example, a pattern could be a special imprint picture, a composed by hand cursive word, a human face, or a talk banner. Given a pattern, its recognition/order may include one of the going with two endeavors (Duda, et. al., 2001):

One of the easiest and most punctual ways to deal with pattern recognition depends on format matching. Matching is a conventional operation in pattern recognition which is utilized to decide the closeness between two substances (points, curves or shapes) of a similar sort. In layout matching, a format (regularly, a 2D shape) or a model of the pattern to be perceived is accessible. The pattern to be perceived is coordinated against the put away layout while considering all passable posture (translation and rotation) and scale changes. The closeness measure, regularly a relationship, might be advanced dependent on the accessible preparing set. Frequently, the layout itself is found out from the preparation set. Format matching is computationally requesting, however the accessibility of quicker processors has now made this methodology increasingly plausible. The inflexible layout matching referenced above, while powerful in some application areas, has various disservices. For example, it would fall flat if the patterns are twisted because of the imaging procedure, perspective change, or substantial intra-class varieties among the patterns (Theodoridis and Koutroumbas, 2003).

In the statistical methodology, each pattern is spoken to regarding d features or measurements and is seen as a point in a d dimensional space. The objective is to pick those features that permit pattern vectors having a place with various classifications to possess smaller and disjoint districts in a d-dimensional element space. The viability of the portrayal space (highlight set) is dictated by how well patterns from various classes can be isolated. Given a lot of preparing patterns from each class, the objective is to set up choice limits in the element space which separate patterns having a place with various classes. In the statistical choice theoretic methodology, the choice limits are controlled by the likelihood circulations of the patterns having a place with each class, which should either be determined or learned. One can likewise adopt a discriminant examination based strategy to classification: First a parametric type of the choice limit (e.g., linear or quadratic) is indicated; at that point the "best" choice limit of the predefined structure is discovered dependent on the classification of preparing patterns. Such limits can be developed utilizing, for instance, a mean squared mistake rule. Statistical pattern recognition is appeared in the accompanying Fig1 (Theodoridis and Koutroumbas, 2003).

Statistical pattern recognition strategy, the classification algorithms dependent on statistical examination. Patterns having a place with a similar class, characterized as a statistically have comparable attributes. In this strategy, a property trademark measurements are portrayed as instances of info pattern is expelled. Each pattern is characterized by a feature vector. By and large, the choice by the classifier, and spotlights on the significance of classification methods. Classifier plan, measurements and pattern information can be handled like to consolidate the probabilities are based. Accordingly, classification, input data space dependent on the assessed likelihood thickness elements of a statistical structure. Different strategies are used to plan a classifier in statistical pattern recognition, contingent upon the sort of information accessible about the class-restrictive densities. A. Dimensionality Reduction: There are two principle motivations to keep the dimensionality of the pattern portrayal (i.e., the quantity of features) as little as would be prudent: estimation cost and classification exactness. A restricted yet notable feature set improves both the pattern portrayal and the classifiers that are based on the chose portrayal. Subsequently, the subsequent classifier will be quicker and will utilize less memory. B. Feature Extraction: Feature extraction methods decide a suitable subspace of dimensionality m (either in a linear or a nonlinear way) in the first feature space of dimensionality d (m=d). Linear changes, for example, central part investigation, factor examination, linear discriminant examination, and projection interest have been generally utilized in pattern recognition for feature extraction and dimensionality reduction The best known linear feature extractor is the primary segment investigation (PCA). iii. Feature Selection: The issue of feature determination is characterized as pursues: given a lot of d features, select a subset of size m that prompts the littlest classification mistake. In the event that huge number of features experienced in the accompanying circumstances:

2) incorporation of multiple data models: sensor data can be displayed utilizing diverse methodologies, where the model parameters fill in as features, and the parameters from various models can be pooled to yield a high-dimensional feature vector (Webb, 2002).

In numerous recognition issues including complex patterns, it is progressively suitable to embrace a various leveled point of view where a pattern is seen as being made out of basic sub patterns which are themselves worked from yet more straightforward sub patterns. The easiest/rudimentary sub patterns to be perceived are called natives and the given complex pattern is spoken to as far as the interrelationships between these natives. In syntactic pattern recognition, a formal relationship is drawn between the structure of patterns and the sentence structure of a language. The patterns are seen as sentences having a place with a language, natives are seen as the letters in order of the language, and the sentences are created by a punctuation. In this way, a huge accumulation of complex patterns can be depicted by few natives and linguistic standards. The syntax for each pattern class must be induced from the accessible preparing tests. Auxiliary pattern recognition is instinctively engaging in light of the fact that, notwithstanding classification, this approach additionally gives a depiction of how the given pattern is developed from the natives. This worldview has been utilized in circumstances where the patterns have a distinct structure which can be caught regarding a lot of tenets, for example, waveforms, finished pictures, and shape examination of forms. The execution of a syntactic approach, nonetheless, prompts numerous challenges which basically have to do with the division of boisterous patterns (to distinguish the natives) and the surmising of the syntax from preparing data. The ascribed language structures which binds together syntactic and statistical pattern recognition. The syntactic approach may yield a combinatorial blast of potential outcomes to be explored, requesting expansive preparing sets and vast computational endeavors. Auxiliary (geometrical, the standard succession) in the pattern recognition approach, a given pattern, formal structure to characterize the essential trademark is diminished. More often than not, the numerical estimations of the set features not just information separated from patterns. Associated with one another or the shared connection between the features, to recognize and classify information to encourage major auxiliary features. At the end of the day, the pattern got from the crude condition of enlightening recognizable proof with formal linguistic structure or language structure from their union basic sub-patterns to characterize complex progressive patterns. Auxiliary technique, each pattern, the segments are treated as an organization (Webb, 2002).

The pattern recognition approaches talked about so far depend on direct calculation through machines. Direct calculations depend on math-related methods. The neural approach applies organic ideas to machines to perceive patterns. The result of this exertion is development of counterfeit neural networks. A neural system is an information handling framework. It comprises of enormous straightforward preparing units with a high level of interconnection between every unit. The handling units work helpfully with one another and accomplish enormous parallel conveyed preparing. The structure and capacity of neural networks reproduce some usefulness of natural minds and neural frameworks. The upsides of neural networks are their versatile learning, self-association and adaptation to non-critical failure capacities. For these exceptional capacities, neural networks are utilized for pattern recognition applications. The absolute best neural models are back-engendering, high-request nets, time defer neural networks and intermittent nets. Typically, just feed-forward networks are utilized for pattern recognition. Feed-forward implies that there is no criticism to the info. Like the manner in which that individuals gain from slip-ups, neural networks likewise could gain from their errors by offering criticism to the info patterns. This sort of criticism would be utilized to remake the info patterns and make them free from error; in this way expanding the execution of the neural networks. Obviously, it is exceptionally unpredictable to develop such sorts of neural networks. These sorts of networks are called as auto cooperative neural networks. As the name suggests, they use back-spread algorithms. One of the primary issues related with back-proliferation algorithms is neighborhood minima. Also, neural networks have issues related with learning speed, engineering choice, feature portrayal, particularity and scaling. In spite of the fact that there are issues and challenges, the

Machine learning is a worldview that may allude to learning from past involvement (which for this situation is past data) to enhance future execution. The sole focal point of this field is programmed learning methods. Learning alludes to change or enhancement of calculation dependent on past "encounters" consequently with no outside help from human. While planning a machine (a product framework), the software engineer dependably has a particular reason at the top of the priority list. For example, think about J. K. Rowling's Harry Potter Series and Robert Galbraith's Cormoran Strike Series. To affirm the case that it was surely Rowling who had composed those books under the name Galbraith, two specialists were locked in by The London Sunday Times and utilizing Forensic Machine Learning they could demonstrate that the case was valid. They build up a machine learning calculation and "prepared" it with Rowling's just as different essayists composing guides to look for and get familiar with the basic patterns and after that "test" the books by Galbraith. The calculation presumed that Rowling's and Galbraith's composition coordinated the most in a few angles. So as opposed to structuring a calculation to address the issue specifically, utilizing Machine Learning, an analyst look for an approach through which the machine, i.e., the calculation will concoct its own answer dependent on the precedent or preparing data set gave to it at first (Duda, et. al., 2001).

• MACHINE LEARNING: INTERSECTION OF STATISTICS AND COMPUTER SCIENCE

Machine Learning was the sensational result when Computer Science and Statistics united. Software engineering centers around building machines that take care of specific issues, and endeavors to recognize whether issues are resolvable by any stretch of the imagination. The primary approach that Statistics generally utilizes is data derivation, demonstrating speculates and estimating unwavering quality of the ends. The characterizing thought of Machine Learning is somewhat unique however mostly subject to both in any case. While Computer Science focus on physically programming PCs, ML tends to the issue of getting PCs to re-program themselves at whatever point presented to new data dependent on some underlying learning strategies gave. Then again, Statistics centers around data deduction and likelihood, Machine Learning incorporates extra worries about the achievability and viability of designs and algorithms to process those data, exacerbating a few learning assignments into a minimal one and execution measures. A third research region firmly identified with Machine Learning is the investigation of human and creature cerebrum in Neuroscience, Psychology, and related fields. The analysts recommended that how a machine could gain for a fact most likely would not be altogether not quite the same as how a creature or a human personality learn with time and experience. Be that as it may, the exploration focused on taking care of machine learning issues utilizing learning methods of human mind did not yield much encouraging outcome so far than the inquires about worried about statistical - computational approach. This might be a result of the manner in which that human or animal mind investigate remains not totally sensible to date. In spite of these inconveniences, composed exertion between human learning and machine learning is extending for machine learning is being used to clear up a couple of learning frameworks finding in human or animals. For example, machine learning system for common refinement was proposed to elucidate neural banners in animal learning. It is truly expected that this joint exertion is to grow broadly in coming years (Duda, et. al., 2001).

Before long, these three requests are so joined and covering that it's about to draw a breaking point or chain of significance among the three. To put it toward the day's end, these three fields are agreeably related and a blend of these methodologies may be used as a technique to make continuously compelling and tricky yields. By and large, Data mining is basically about decoding any kind of information, anyway it builds up the system for both man-made thinking and machine learning. Eventually, it test data from various sources just as it examinations and sees pattern and connections that exists in those data that would have been difficult to decipher physically. Along these lines, information mining is surely not a unimportant methodology to exhibit a hypothesis yet procedure for outline appropriate speculations. That mined data and the comparing patterns and theories might be used the reason for both machine learning and man-made brainpower. Computerized reasoning might be comprehensively characterized as machines those being able to tackle a given issue individually with no human mediation. The solutions are not customized specifically into the framework but rather the important data and the AI translating that data produce an answer independent from anyone else. The elucidation that goes underneath is only a data mining calculation. Machine learning adopts elevate the strategy to a propelled dimension by giving the data basic to a machine to prepare and adjust reasonably when presented to new data. This is known as "preparing". It centers around

statistical measures to enhance its capacity to translate new data and produce increasingly viable outcomes. Obviously, a few parameters ought to be "tuned" at the beginning dimension for better efficiency. Machine learning is the a dependable balance of computerized reasoning. It is doubtful to design any machine having capacities related with knowledge, similar to language or vision, to arrive without a moment's delay. That undertaking would have been practically difficult to understand. Additionally, a framework can not be considered totally savvy on the off chance that it came up short on the capacity to take in and enhance from its past exposures.

A mind-boggling number of ML estimation have been designed and displayed over past years. Only one out of every odd individual of them are commonly known. Some of them didn't satisfy or tackle the issue, so another was exhibited in its place. Here the algorithms are broadly amassed into two class and those two social events are further sub-disengaged. This region attempt to name most conspicuous ML algorithms and the accompanying portion investigates three most comprehensively used ML algorithms.

1. Supervised learning — Input data or preparing data has a pre-decided mark for example Genuine/False, Positive/Negative, Spam/Not Spam and so on. A capacity or a classifier is constructed and prepared to foresee the mark of test data. The classifier is appropriately tuned (parameter esteems are adjusted) to accomplish a reasonable dimension of accuracy. 2. Unsupervised learning - Input data or preparing data isn't marked. A classifier is designed by reasoning existing patterns or bunch in the preparation datasets (Webb, 2002). 3. Semi-supervised learning - Training dataset contains both marked and unlabelled data. The classifiers train to get familiar with the patterns to classify and mark the data just as to anticipate. 4. Reinforcement learning - The calculation is prepared to delineate to circumstance with the goal that the reward or criticism flag is augmented. The classifier isn't customized straightforwardly to pick the activity, yet 5. Transduction - Though it has comparative qualities with regulate learning, however it doesn't build up an express classifier. It endeavors to foresee the yield dependent on preparing data, preparing mark, and test data. 6. Learning to learn - The classifier is prepared to gain from the predisposition it prompted amid past stages. 7. It is essential and proficient to sort out the ML algorithms as for learning methods when one have to think about the noteworthiness of the preparation data and pick the classification decide that give the more prominent dimension of accuracy.

Regression investigation is a piece of prescient examination and adventures the co-connection between ward (target) and free factors. The striking relapse models are:Linear Regression, Logistic Regression, Stepwise Regression , Ordinary Least Squares Regression (OLSR), Multivariate Adaptive Regression Splines (MARS) , Locally Estimated Scatterplot Smoothing (LOESS, etc.

Occasion based or memory-based learning model stores examples of preparing data as opposed to building up an exact meaning of target work. At whatever point another issue or precedent is experienced, it is inspected as per the put away occasions so as to decide or anticipate the objective capacity value. It can just supplant a put away occurrence by another one if that is a superior fit than the previous. Because of this, they are otherwise called victor take-all strategy. Examples: K-Nearest Neighbor (KNN), Learning Vector Quantization (LVQ), Self-Organizing Map (SOM), Locally Weighted Learning (LWL) and so forth.

Regularization is just the way toward checking over fitting or lessen the exceptions. Regularization is only a straightforward yet amazing alteration that is enlarged with other existing ML models ordinarily Regressive Models. It smoothes up the regression line by rebuking any twisted of the bend that attempts to coordinate the anomalies. Examples: Ridge Regression, Least Absolute Shrinkage and

A choice tree builds a tree like structure including of conceivable solutions to an issue dependent on specific limitations. It is so named for it starts with a solitary straightforward choice or root, which at that point forks off into a number of branches until a choice or prediction is made, framing a tree. They are favored for its capacity to formalize the issue close by procedure that thusly helps distinguishing potential solutions quicker and more precisely than others. Examples: Classification and Regression Tree (CART), Iterative Dichotomiser 3 (ID3), C4.5 and C5.0, Chi-squared Automatic Interaction Detection (CHAID), Decision Stump, M5, Conditional Decision Trees and so on (Webb, 2002).

A gathering of ML algorithms utilize Bayes' Theorem to take care of classification and regression issues. Examples: Naive Bayes, Gaussian Naive Bayes, Multinomial Naive Bayes, Averaged One-Dependence Estimators (AODE), Bayesian Belief Network (BBN), Bayesian Network (BN) and so forth.

SVM is so famous a ML procedure that it tends to be its very own gathering. Ituses an isolating hyper plane or a choice plane to demarcate choice limits among a lot of data points classified with various marks. It is an entirely supervised classification calculation. At the end of the day, the calculation builds up an ideal hyper plane using input data or preparing data and this choice plane in turns categories new examples. In light of the portion being used, SVM can perform both linear and nonlinear classification.

Clustering is worried about utilizing instilled pattern in datasets to classify and mark the data accordingly. Examples: K-Means, K-Medians, Affinity Propagation, Spectral Clustering, Ward hierarchical clustering, Agglomerative clustering. DBSCAN, Gaussian Mixtures, Birch, Mean Shift, Expectation Maximization (EM) and so on.

Algorithms Association rules help find connection between's obviously unassociated data. They are generally utilized by online business sites to anticipate client practices and future needs to elevate certain engaging items to him. Examples: Apriori calculation, Eclat calculation and so on. A model dependent on the constructed and operations of real neural networks of people or creatures. ANNs are viewed as non-linear models as it endeavors to find complex relationship among information and yield data. In any case, it draws test from data instead of thinking about the whole set and subsequently decreasing expense and time. Examples: Perceptron, Back Propagation, Hop-field Network, Radial Basis Function Network (RBFN) and so forth.

These are progressively modernized forms of ANNs that profit by the abundant supply of data today. They are utilizes larger neural networks to take care of semi-supervised issues where real segment of a flourish data is unlabeled or not classified. Examples: Deep Boltzmann Machine (DBM), Deep Belief Networks (DBN), Convolutional Neural Network (CNN), Stacked Auto-Encoders and so forth (Duda, et. al., 2001).

Dr. D. Ashok Kumar and N. Kannathasan [5] studied utility of data mining and pattern recognition strategies for soil data mining and its unified regions. The suggestions emerging from this exploration review are: A correlation of various data mining strategies could create a proficient calculation for soil classification for multiple classes. The advantages of a more noteworthy comprehension of soils could enhance profitability in cultivating, look after biodiversity, diminish dependence on manures and make a superior incorporated soil the executives framework for both the private and open parts. Farah Khan and Dr. Divakar Singh [6] try to give a diagram of some past inquires about and contemplates done toward applying data mining and explicitly, affiliation rule mining systems in the rural space. They have additionally attempted to assess the present status and conceivable future patterns around there. The hypotheses behind data mining and affiliation rules are exhibited toward the start and an overview of various procedures connected is given as a component of the advancement. Amina Khatra [7] demonstrated that utilizing shading based picture division it is conceivable to separate the yellow rust from the wheat crop pictures. Further, the portioned yellow rust pictures can be presented to estimation calculation where the genuine entrance of the yellow rust might be assessed in the yield. This sort of picture division might be utilized for mapping the adjustments in land utilize land spread assumed control transient

situating of the cameras introduced so as to gain the pictures from the field. Archana A. Chaugule and Dr. Suresh Mali [8] in their examination Shape-n-Color feature set beat in practically every one of the occurrences of classificationfour Paddy (Rice) grains, viz. Karjat-6, Ratnagiri-2, Ratnagiri-4 and Ratnagiri-24. They utilized Pattern classification was finished utilizing a Two-layer (for example one-shrouded layer) back proliferation supervised neural networks with a solitary concealed layer of 20 neurons with LM preparing functions.The fifty-three features were utilized as contributions to a neural system and the sort of the seed as target. Abirami et al. [9] utilized canny edge recognition, thersolding and scaled conjugate slope preparing with 9 neurons in shrouded layer for evaluating basmati rice granules. Scaled Conjugate Gradient Training based Neural Network could classify granules with the accuracy of 98.7%. Different reviewing frameworks have been created, which utilize distinctive morphological features for the classification of various oat grains.

In spite of the fact that different researchers have added to ML and various algorithms and procedures have been presented as referenced before, on the off chance that it is firmly considered the greater part of the down to earth ML approach incorporates three primary supervised calculation or their variation. These three are in particular, Naive Bayes, Support Vector Machine and Decision Tree. Dominant part of researchers have used the idea of these three, be it specifically or with a boosting calculation to improve the productivity further. These three algorithms are examined quickly in the accompanying segment.

It is a supervised classification methoddeveloped utilizing Bayes' Theoremof restrictive likelihood with a 'Guileless' presumption that each pair of feature is commonly autonomous. That is, in less difficult words, nearness of a feature isn't affected by nearness of another using any and all means. Regardless of this over-disentangled presumption, NB classifiers performed great in numerous handy circumstances, as in content classification and spam detection. Just a little measure of preparing data is needto gauge certain parameters. Adjacent to, NB

Another supervised classification calculation proposed by Vapnik in 1960s have as of late pulled in a noteworthy consideration of researchers.The basic geometrical clarification of this approach includes determiningan ideal isolating plane or hyperplane that isolates the two classes or groups of data points evenhandedly and is equidistant from them two. SVMwasdefinedat first for linear dissemination of data points. Afterward, the part work was acquainted with tackle nonlinear datas also.

A classification tree, prominently known as choice tree is a standout amongst the best supervised learning calculation. It develops a chart or tree that utilizes branching method to show each plausible result of a choice. In a choice tree portrayal, each interior hub tests a feature, each branch relates to result of the parent hub and each leaf at long last relegates the class mark. To classify an occurrence, a best down approach is connected beginning at the foundation of the tree. For a specific feature or hub, the branch agreeing to the estimation of the data point for that property is considered till a leaf is come to or a name is chosen. Presently, the exhibitions of these three were generally contrasted utilizing a lot of tweets and marks positive, negative and unbiased. The crude tweets were taken from Sentiment 140 data set. At that point those are pre-handled and named utilizing a python program. Every one of these classifier were presented to same data. Same calculation of feature determination, dimensionality reduction and k-overlay approval were utilized for every situation. The algorithms were thought about dependent on the preparation time, prediction time and accuracy of the prediction. The test result is given below:

Naïve Bayes (Gaussian) 2.708 0.328 0.692

SVM 6.485 2.054 0.6565 Decision Tree 454.609 0.063 0.69 The system architecture of the OS-ELM executed in this examination comprises of 1056 information yield of 0 demonstrates a favorable example and a yield of 1 shows a threatening example. In the OS-ELM just a single parameter should be resolved, which is the number of concealed layer neurons ), since the OS-ELM is a SLFN. The technique to look for the ideal number of the shrouded layer neurons ) in the OS-ELM is proposed by (Huang et al., 2004), which shows that the number of concealed neurons, shift in the range from 20 to 200.

Both SVMs and ANNs are considered as discovery demonstrating systems. Albeit the two algorithms share a similar structure, yet the learning methods for the two algorithms are totally unique. ANNs attempt to limit the preparation error, though SVMs decrease limit utilizing the SRM guideline. Examination results of BPNN and OS-ELM as opposed to the SVM based model are organized in Table 6.4, which are acquired for testing the 70 tests from the nearby dataset (UMMC and MIAS). The trial results in Table 6.4 demonstrate that the SVM based approach beats the BPNN and the OS-ELM concerning the general classification accuracy. This is on the grounds that the ideal results for paired classification are gotten by the SVM based model, where parameters: sensitivity, specificity, TPF, FPF and MN are in ideal reaches.

VII. CONCLUSION

A Paper on the pattern recognition has been exhibited. It has been appeared incredible methods exist, in any case, care must be taken to assemble vigorous and predictable classifiers. The best approach for the unpracticed client is by all accounts the utilization of classical statistical apparatuses, since fitting and play works for this situation. Pattern recognition should be possible both in typical PCs and neural networks. PCs utilize regular number juggling algorithms to recognize whether the given pattern coordinates a current one. It is a clear technique. It will say either yes or no. It doesn't endure uproarious patterns. Then again, neural networks can endure clamor and, whenever prepared appropriately, will react accurately for obscure patterns. Neural networks may not perform supernatural occurrences, yet whenever built with the best possible architecture and prepared accurately with great data, they will give astounding results, in pattern recognition as well as in other logical and business applications. The foremost target of ML researchers is to design increasingly effective (regarding both time and space) and down to earth universally useful learning methods that can perform better over a far reaching area. With regards to ML, the effectiveness with which a technique uses data assets that is likewise an imperative execution worldview alongside existence intricacy. Higher accuracy of prediction and humanly interpretable prediction rules are likewise of high significance. ML gives a product the adaptability and flexibility when vital. Notwithstanding some application (e.g., to compose network multiplication programs) where ML may neglect to be valuable, with increment of data assets and expanding request in customized customisable programming, ML will flourish in not so distant future. Other than programming advancement, ML will presumably however help change the general standpoint of Computer Science. By changing the characterizing question from "how to program a PC" to "how to empowerit to program itself," ML cloisters the

simply preparing it. Moreover, it will help change Statistical guidelines, by giving increasingly computational position. Clearly, the two Statistics and Computer Science will likewise adorn ML as they create and contribute further developed hypotheses to adjust the method for learning.

1. R. O. Duda, P. Hart and D. Stork (2001). Pattern Recognition, USA: John Wiley & Sons, 2001. 2. S. Theodoridis and K. Koutroumbas (2003). Pattern Recognition, USA: Academic Press, 2003. 3. A. Webb (2002). Statistical Pattern Recognition, England: John Wiley & Sons Ltd., 2002. 4. R. O. Duda, P. Hart and D. Stork (2001). Pattern Recognition, USA: John Wiley & Sons, 2001. 5. S. Theodoridis and K. Koutroumbas (2003). Pattern Recognition, USA: Academic Press, 2003. 6. A. Webb (2002). Statistical Pattern Recognition, England: John Wiley & Sons Ltd. 7. C. M. Bishop (2006). Pattern Recognition and Machine Learning, Singapore: Springer Science+Business Media, LLC, 2006 8. D. S. Gunal (2008). "AUTOMATED CATEGORIZATION SCHEME FOR DIGITAL LIBRARIES IN DISTANCE LEARNING:A Pattern Recognition Approach," Turkish Online Journal of Distance Education-TOJDE, vol. 9, p. Number:4 Article 1, Octomber 2008. 9. M. Steenweg, A. Vanderver, S. Blaser, T. d. K. A Bizzi, G. Mancini and B. F. W. N. v. d. K. M. van Wieringen W.N. (2013). "Magnetic resonance imaging pattern recognition in hypomyelinating disorders.," p. 136(Pt 9):2923, Sep 2013.

mane.paridnya39@gmail.com