You are on page 1of 9


Bijay Baran Pal1, Somsubhra Gupta2, Debashis Sanki3


Professor & Head, Department of Mathematics, University of Kalyani, Kalyani, West Bengal -741235 2 Asst. Professor, Department of Information Technology, JIS College of Engineering Kalyani, Nadia, West Bengal, 741235, INDIA 3 Lecturer, Department of Information Technology, JIS College of Engineering Kalyani, Nadia, West Bengal, 741235, INDIA 1,,

Objective of the system is to measure the effect of sound wave / music sample(s) on human brain with a sense of optimization and to discover some interesting measure from the recorded image base based upon the metadata of music sample and to relate them with the physical feedback given by the person who is under testing. The optimization is to be achieved using Genetic Algorithm principle. This dynamical system will initiate work by means of generate and test with a large population of human being and number of variant music sample effectively yielding large data warehouse. Various trends will be analyzed and discovered from this data warehouse by means of Data Mining techniques. Once this correspondence is established with a reasonable degree of aspiration from the data warehouse, it will be applied to examine how much accurately the association rule is working for various case studies. The major motivation behind this work is the earlier work to develop cognitive system that can treat patient of non surgical category (both psychological and non-psychological) using Music Therapy treatment convention. The successful application of it may well open the scope of utilizing music as an effective tool in various human needs viz. relaxation, temperament restoration and above all therapeutic needs. Keywords: Neural Network (NN), Knowledge Discovery (KD), Magnetic Resonance Imaging (MRI), Brain Model, and Music Impact.

The entire project will have four phases. Each phase is discussed illustrating pictorial representation of system flow. The assumption is that any music sample is a collection of many discrete frames. Each trial should ideally get single frame at a time. But by means of lowing down the fractional time we can address number of frames in a small span thus incorporating continuity in a connectionist network (to emulate the kind of structure brain possesses). The change of states of human brain will be dynamically tracked by Biomedical Instrumentation process (MRI / CT scan) through retrieval of several brain images in small time. In view of measuring effects on human brain it is essential to be informed about it. The human nervous system is built of cells called neurons. This system is highly complex. About 1011 neurons participate in perhaps 1015 interconnections over transmission paths that may range for a meter or more. Each neuron shares many characteristics with the other cells in the body but has unique capabilities to receive, process and transmit electrochemical signals over the neural pathways that comprise the brains communication system. Dendrites extend from the cell body to other neurons where they receive signals at a connection point called a synapse. On the receiving side of the synapse, these inputs are conducted to the cell body. There they are summed, some inputs tending to excite the cell, others tending to inhibit its firing. When the cumulative excitation in the cell body exceeds a threshold, the cell fires by sending a signal down the axon to other neurons. This basic functional outline has many complexities and exceptions; nevertheless, most artificial neural networks model only these simple characteristics. Following are some facts about human brain The human nervous system is built of cells called neurons. This system is highly complex. 11 10 neurons

Very high connectivity. Any neuron can be connected to about 105 other neurons.

1014 1015 synapses. That is, inter-connections between neurons. Activity is highly parallel. Auto-associative memory Based upon an idea about human brain discussed above we proceed with several trials. A particular trial incorporates a specific music sample and examines its effect on a human being. Any trial consists of the following steps. The Bio-medical Instrumentation Process (E.g. MRI) or Biological Connectionist Neural Network process is elaborately discussed later.

2. Initiation of Brain-Image base

First take a snap of image or sequence of images moment before the music sample is initialized. Send those images to the image base properly enlisting the time and other environmental inputs sans music sample. Following is the illustration of it through diagram Initial state of human brain Scan it MRI / CT Scan

Image base

(Image or sequence of images moment before initializing the music sample/ later on various frames)

2.1 Acceptance of sound frames and Induction Process

Next the music sample is initiated. It can be stored using any format (standard format like wav / mp3). The system under discussion can have its own music player or any standard music player can be embedded. The first is advantageous as the metadata is to be stored in order to maintain a record corresponding to played music file which itself is an input. This means detail about this music sample will be stored. So, well have a record about the played music comprising the Music domain, type of the Music (Traditional / classical/ pop/ folk etc.), classification (vocal / instrumental), tone (mono/ poly), key nodes and more fields so as to get the background detail of the music.

Music Sample (mp3 / wav / else.

(Sample music stored in Binary or any other format)

System plays
(System may possesses in built sound player or standard media player can be embedded) High- speed parallel distributed connectionist Architecture (and Process)

Sound wave in various discrete frames

Human brain

Environmental inputs and other factors. The above two steps will be iterated as long the music sample generates sound wave and brain images generated in sequence are stored in the image-base.

Iteration span should be so small that very small incremental portion of the sample music can be contained in frames thus resulting numerous images in a small span. The generation of sound, which otherwise seemed to be a continuous event, should be tracked in so small time span through iteration that several images can be retrieved to measure even smaller change over. The above will improve analysis since very small effect or transitional changes can be tested through numerous images.

2.2 Connectionist Network and Brain images generation

The next step illustrates generalization through connectionist network. This means how the Human brain, which itself is a connectionist network, is induced continuously through various frames that the sample music comprised of. Subsequently various brain images are scanned by Biomedical Instrumentation Process within fractional time span thus almost ensuring a dynamical productions of brain images those will be kept in the image base. A correspondence will be established between the images and the frames based upon which these images are produced. It is to be reminded that we have the records corresponding to each partitioned frames that the music modules consists of. Now well illustrate INDUCTION PROCESS of brain through sound wave at various time instances and the consequent diagram illustrate how the dynamical system retrieves brain images frequently and stores them in the image-base.


The number of inputs at various time instances.

In the above diagram Ti s refers the time instance of inputs. The inputs correspond to various pitch frequencies, contained in the frame, constituting the music module. Following is the Connectionist Network emulating the way brain gets the input signal. Consequently the MRI retrieves the images. Later on we have discusses elaborately on MRI in a separate chapter. Brain image X1 X2 XN against Frame 1 T1


BI 1



BI 2

Hidden layer TN Frame3

BI 3

Cumulative induction of hidden layer in a multi-layer induction process.

These are stored in Image base

In the above X1, X2 XN are input signals. In the hidden layer the NET impulse those are being carried by the synaptic connections represented by line in various intermediate (hidden layer) stations are represented by and corresponding snapshots of brain images against the frames are stored in image base. Furthermore we can have a comprehensive measurement of the carried impulse by the Human brain in the induction process. This can be expressed as a function of number of inputs at various points of time and synaptic strength between various biological neurons and the neighboring neurons that they induce. The first order characteristics of

biological neurons are analyzed. A set of inputs x1, x2,.., xn is considered. These inputs collectively referred to as the vector X, corresponds to the signals into the synapses of a biological neurons. Each signal is multiplied by an associated weight w1, w2,., wn before it is applied to summation block, . Each weight corresponds to the strength of a single biological synaptic connection. The set of weights is referred to collectively as the vector W. The summation block, corresponding roughly to the biological cell body, adds all of the weighted inputs algebraically producing an output NET. This is NET = x1w1 + x2w2 + x3w3 ++ x n w n = x i w i This may be compactly stated in vector notation as NET = XW It is then compared to the threshold . This actually indicates whether the induction is up to the desired level. The NET signal is further processed by an activation function F to produce neurons output signal OUT. OUT = F (NET - ) For constant threshold value , =1 if NET> OUT =0 if NET< This indicates whether a neuron can fire an adjacent neuron to induce or not.

2.3 Analysis of Brain Images and Optimization

Here we apply pattern recognition technique for processing brain images, which are snapped by the dynamical system corresponding to various frames of the music sample. The objective here is to analyze the images, pointing out changeover with respect to change of frames and furthermore to establish a mapping between the music characteristics carried by frames with the brain images. Here we can use any standard pattern recognition technique. These trends will be stored in a data warehouse from which association rule can be derived to analyze plausible changeover of a human being through sound effect. It is generally accepted that audio frequencies above 20 KHz dont affect human sensory perceptron since they are beyond the audible range. The question of whether the existence of such inaudible high-frequency components may affect the acoustic perception of audible sounds remains unanswered. In this study, noninvasive physiological measurements of brain responses are used to provide evidence that sounds containing high-frequency components (HFCs) above the audible range significantly affect the brain activity of listeners. An audible low-frequency component (LFC) (below 22 kHz and an HFC above 22 kHz) Brain electrical activity and regional cerebral blood flow (rCBF) were measured as markers of neuronal activity while subjects were exposed to sounds with various combinations of LFCs and HFCs. None of the subjects recognized the HFC as sound when it was presented alone. Nevertheless, the power spectra of the alpha frequency range of the spontaneous electroencephalogram (alpha-EEG) recorded from the occipital region increased with statistical significance when the subjects were exposed to sound containing both an HFC and an LFC, compared with an otherwise identical sound from which the HFC was removed (i.e., LFC alone). In contrast, compared with the baseline, no enhancement of alpha-EEG was evident when either an HFC or an LFC was presented separately. Positron emission tomography measurements revealed that, when an HFC and an LFC were presented together, the rCBF in the brain stem and the left thalamus increased significantly compared with a sound lacking the HFC above 22 kHz but that was otherwise identical. Simultaneous EEG measurements showed that the power of occipital alphaEEGs correlated significantly with the rCBF in the left thalamus. Psychological evaluation indicated that the subjects felt the sound containing an HFC to be more pleasant than the same sound lacking an HFC. These results suggest the existence of a previously unrecognized response to complex sound containing particular types of high frequencies above the audible range. This phenomenon is termed as hypersonic effect. At the time of deriving interesting measure from recorded effect, one has to satisfy through repeated training result that result outcomes incorporate less non uniformity. Because if various events of test results are not uniform then their remains a high risk of being terminated with triviality. In such occasion even local maximum cant be reached. The best way to achieve the optimization is to take a initial population with large variation in characteristics. All the initial measurements are stored. Then these are utilized in the next iteration with similar kind of sound wave before concluding the interesting measure through association rules. Since the human characteristics are intrinsically fuzzy in nature, the statistical optimization is hardly acceptable. The best way to test it is through successive iterations incorporating genetic algorithm principle by which the process continues until an aspiration level is reached. This aspiration level may very much be a range rather than a fixed point because predefined threshold is unlikely to be satisfied. Rather an auto associative achievement of aspiration level is more acceptable. This conclusion is dominated by the fact that the same human being may not be affected with same scale at always

by the same music module or sound wave of certain frequency. This is due to variation in environmental input which is highly non-deterministic in nature.

3. Framework in Data Warehouse

It comprised of various relational database. We discuss them one after another depicting pictorially the frame and concluding with the collective shape of the Data Warehouse where well later try to establish various association to derive rules during analyzing the trends. First consider the Music Sample Database. As we require repeating the trial described in the previous sections using a huge variation of music sample on population with large size in order to analyze trend, it is increasingly required to have elaborate background detail correspond to each of the music samples that are being used for testing purpose. However it is not essential to test effect of all the music samples to the entire person in the population. It is rather constrained by means of background details available in the relational database. However testing all the samples for total population makes association rule more robust amidst cost complexity.

Music Sample database

Fields Music ID Record 1 .. Record 2 .. And so on. And preferably many other fields robust analysis. Type Classification .. .. Tone .. .. Pitch Principal Nodes . .

including singer, if it be vocal. The more background detail will only facilitate

Person database
Field Person ID Age-group Financial Background Professional Background with other fields of relevance. The idea is to derive later on interesting patterns by establishing association between persons and music samples. The brain image analysis and physical feedback of the person will enhance the acceptability of the knowledge that is being mined from the Data Warehouse. Following is the Master Database comprising of music sample, persons and the effect that is been measured through pattern recognition and image analysis of human brain images and also through collective physical feedback by the human being. We have assumed a population size N. Music variation M. #Trial No Measured Effect through Physical brain image analysis feedback 1 M1 P1 e11** f11** 2 M1 P2 e12 f12 .. N+1 M2 P1 e21 f21 N+2 M2 P2 e22 f22 ** eij s and fij s contain the resultant effect of brain images and physical feedback respectively of the music sample i on the person j. Ideally we should express effect using eijk instead of eij as we have observed that the effect of a sample music not always same on a particular person at various points of time. It depends on various environmental inputs also. eijk stands for effect of sample music i on person j at kth trial. Music Sample Person ID Regional Ethnic Background

3.1 Knowledge Discovery in Database

From the discussion in the last section it is quite evident that now we can embed two databases, one corresponding to the person and another corresponding to the Music sample, to the master database for establishing direct link with the measured effect through brain imaging or physical feedback. This will enable an enhanced view of the entire. Consequently we can have the following clusters of fields available in the large Data Warehouse. Field Measured Effect Attributes such as Abstract Lyric Aggressive drumming . Etc. Physical feedback Attitude Behavior Concentration Fitness Music Sample Music ID Type Classification Tone Pitch Principal Nodes Person Person Id Age group Regional Educational Financial Professional Environmental Inputs Other fields of relevance

Sub field

Now we can make an exhaustive searching in the Data Warehouse to find interesting measure to analyze trend, viz. Which age group prefers e.g. Indian Classical Ragas? What are the trends in the other fields corresponding to it? i.e. people with which educational, regional and professional background prefers it most. Whether at all those background have any effect on the choice. What is the general trend among the young field? Whether at all that varies with regional background or any other. Whether certain tune / Pitch have attracts people with certain age group. A sample-mined trend can be like the following X, educational status (X) = post graduate financial status(X) = upper middle age-range (X)=32..37 choice = classical. (Occurrence: 45%) This signifies 45% of the population having the background detail mention above prefers Classical music. We will concentrate on Constrain based Association Mining (otherwise it will be very tedious or voluminous). For given set of task relevant data, the data mining process may uncover thousands of rules, many of which are uninteresting to the user. In constraint-based mining, mining is performed under the guidance of various kinds of constraints provided by the user. These constraints include the following: Knowledge type constraint: these specify the type of the knowledge to be mined, such as association. Data constraints: These specify the set of task relevant data. Dimension / level constraint: These specify the dimension of the data, or levels of the concept hierarchies, to be used. Interestingness constraints: these specify the thresholds on statistical measures of rule interestingness, such as support and confidence. Rule constraints: these specify the form of rules to be mined. Such constraints may be expressed as meta rules (rule templates), as the maximum /minimum number of predicates that can occur in the rule antecedent or consequent or relationship among the attributes, attribute values and /or aggregates.

4. Magnetic Resonance Imaging (MRI)

The biometric images in the context of the present project refer to anatomical image of human brain by modern diagnostic scanners such as Computer Tomography or Magnetic Resonance Imaging Scanner. Conventional scanning is performed in search of anatomical anomaly in the subject. Functional imaging refers to a series of scans recording changes occurring over time. Over the past few decades several degenerative nervous disorders came in focus. The patients suffering from these diseases do not show any anatomical abnormalities, but the area of brain responsible for cognitive senses, motor activities and balance gradually ceases to function. The existing diagnoses based on anatomical studies fail to detect at early stage when the diseases are curable. Duration at later stage doesnt come to much help as for the patients. For early detection require scanning brain. Why scanning is required? To understand it requires clear idea about functional relationship of the human brain body coordination. Each functioning of the human body is controlled by brain like CPU of computer system. Communication, between brain and different organs, is made by the biological neural systems. Total brain area can be divided into 52 parts, each performs different function at a given instance, and it never does a particular function forever. Brain process is divided into three parts Top level: Computational theory (concern with the goal of the process & strategy required to achieve that goal) Code and Algorithm: Nature of the code and algorithm by which transformation from one code to another is performed. Implementation: by interaction with the neurons. Now when a complex task is given a single part of the brain cant perform it. Total task is divided in to modules and each module is distributed to various parts of the brain. Thus the set of the parts of brain can perform this task that bring the concept of local specialization and global integration. Particularly this part of the project can be considered as a part of the greater effort to map different functions of the human body with corresponding regions in the brain. The aim of this work is identification of different functional part of the brain and hence the active regions of a subject can be compared with those of fitter persons leading to fast non-evasive diagnosis of functional disorder. This enables our objective of checking the authenticity of the physical feedback of the person through the behavioral analysis of the brain images.

Functional magnetic Resonance Imaging sensitive to oxy-hemoglobin concentration are performed on a person over a period of time. The subject, who is under complete physical and mental rest for sometime, performs a single voluntary action like finger tapping / speaking name of a set of flowers for some duration then no reaction for same amount of duration and then again react for same duration and so on, repetitively during the experiment. The variation of oxy-hemoglobin concentration in brain volume is recorded as data and observed for changes or activations in the brain.

4.1 MRI Synopsis

Following steps are sequentially put subject in big magnetic field (leave them there). Transmit radio waves in to subjects [about 3 minutes]. Turn off radio wave transmitter. Receive radio waves re-transmitted with magnetic fields during read-out interval [10-100 ms: MRI is not a snapshot] Store measured radio wave data vs. time Now go to step 2 and repeat for some sizable amount. Process raw data to reconstruct images Allow subject to leave scanner (Optional). The physics of MRI can be divided in to two major parts: Obtaining tissue contrast based on some particular element, such as oxy-hemoglobin concentration. Producing an image in space based on the contrast signal Atoms of different elements within brain tissue (tissue water) have a characteristic spin magnetic moment due to the spinning electrons. Thus each of these atoms can be considered as small magnet. In normal state magnetic moment of the atoms cancel each other. For MR imaging at first the subject (human brain) is placed under a strong magnetic field, typically in the range of 1.5 to 4 Tesla. Depending on the strength of the magnetic field produced, it tries to line up the little magnets with external lines of force. The net magnetic moment of the atom M moves around the applied field and then gradually becomes parallel.

4.2 MRI Precession

Magnetic field causes M to rotate (or process) about the direction of B at a frequency proportional to the size of B-42 million times per second (42MHz), per Tesla of B. If M is not parallel to B, then it proceeds, clockwise, around the direction of B. However normal (fully relaxed) situation has M parallel to B. This means there wont be any precession. No precession of M parallel to B.

5. Case Study
To examine the physiological effect of sounds with an inaudible frequency range, 11 subjects were presented with the FRS, HCS, and baseline conditions. In this experiment, a cutoff frequency of 26 kHz with a steeper cutoff attenuation of 170dB/ octaves was employed to separate HFCs from LFCs. This relatively high cutoff frequency was chosen because when a cutoff frequency lower than 26 kHz is used the skirts of the power spectrum of the filtered HFCs extend below 20 kHz and generate sounds containing components below 20 kHz. It is widely known that the upper limit of the audible range of humans varies considerably. It usually corresponds to around 15 or 16 kHz in young adults and sometimes below 13 kHz in the elderly, and some people can recognize air vibrations of 20 kHz as sound. When a cutoff frequency of 26 kHz is employed with the steeper cutoff attenuation, the power spectrum of the filtered HFCs under 20 kHz falls below the system noise level. Therefore we selected a cutoff frequency of 26 kHz, which is sufficiently high to completely exclude contamination by audible sound components in all of the subjects. In accordance with conventional recordings of background EEG activity, subjects were asked to keep their eyes naturally closed during the experiment to eliminate any effects of visual input. The presentation of the sounds in both FRS and HCS conditions lasted 200 s, which included the entire piece of music. The baseline

condition also lasted 200 s without sound presentation. The inter-session intervals were 10 s. Two recording sessions were repeated for each condition in the following order: baselineFRSHCSFRSHCSbaseline. So by means of analysis of brain images with respect to sound with the above example and case study we can infer that a correspondence can be establish between the input frame and the brain images. The physical feedback from the person under testing can be collectively stored in order to get the idea that how much is the effect of the music sample on the person. This, at later phase of analysis, in the knowledge discovery of database will ease out to establish a relation between the person (with full knowledge about his / her background detail) and the sample music. Basically this will establish a trend on taste of music of persons of various backgrounds.

6. Conclusion
The approach provides a way to keep track of sound wave /music samples impact on human mind. The recorded effect is measured in correspondence with verbal feedback of the human being in a large population with a Neural Network approach of training and test. Then some trends are tried to be analyzed using data mining concepts those are almost equivalent to rule. These rules, because of their tested acceptance, can be applied to various human beings with matching characteristics. This is provided with a cost effective way. From the viewpoint of utility of the approach, in todays world music has some greater significance than ever. Music tunes the heart, controls the mind set by making emotional adjustments. People use to have music even during their course of work to improve efficiency. Moreover it is acceptable tool to relax and restores the sense of normalcy from mind exhaustion. In a word music is a universal language that influences human existence. So in therapeutic needs, the outcome of the present work is very vital.

Agarwal R., and J.C. Shafer, 1996, Parallel mining of association rules?: Design, Implementation, and experience. IEEE Trans. Knowledge and Data Engineering, 8: 962-969. Ankerst M., C. Elsen, M. Ester, and H.P. Kriegel, 1999 Aug., Visual Classification: An Interactive approach to decision tree construction, In Proc. 1999 Int. Conf. Knowledge Discovery and Data Mining (KDD99), San Diego, CA, Pages 392-396. Aoki P.M., Generalizing search in generalized search trees, In Proc. 1998 Int. Conf. Data Engineering (ICDE98), Pages 380-389. Avner S., Discovery of comprehensible symbolic rules in a neural network. In Intl. Symposium on Intelligence on Neural and Biological Systems, pages 64-67, 1995. Berthold M., and D. J. Hand, 1999, Intelligent Data Analysis: An Introduction, Springer-Verlag. Bezdek J.C., and S.K. Pal, 1992,Fuzzy Models for Pattern Recognition: Methods that search for structures in Data. IEEE Press. Breiman L., 1996,Bagging Predictors, Machine Learning, 24: 123-140. Brown D. E., V. Corruble and C.L. Pittard, 1993, A comparison of decision tree classifiers with back-propagation neural networks for multi-modal classification problems, Pattern Recognition, 26: 953-961. Craven M.W., and J.W. Shavlik, 1997, Using neural networks in data mining. Future Generation Computer Systems, 13: 211-229. Gupta S., A. Datta, 2005, Cognitive Systems incorporating Artificial Intelligence: A behavioral model of Music Therapy Treatment Convention, Proceedings of International Conference on Cognitive Systems 2005, NIIT-CRCS, New Delhi. Han J., and M. Kamber, Data Mining Concepts and Technique, Morgan Kaufman. Konar. A, Artificial Intelligence and Soft Computing: A behavioral and cognitive model of Human Brain, CRC Press. Kosko B., Neural Networks and Fuzzy Systems, A dynamical system approach to Machine Intelligence, Prentice Hall of India. Pokrajac D., and A. Lazarevic, 2004, Localized Neural Network based Distributed Learning for Knowledge Discovery in Protein databases, Proceedings of International Joint Conference on Neural Network (IJCNN04), IEEE. Tistarelli M., J. Bigun and A. Jain, Biometric Authentication, Springer Verlag.