Interactive Machine Learning

Transcription

1 Interactive Machine Learning Jerry Alan Fail, Dan R. Olen, Jr. Computer Science Department Brigham Young Univerity Provo, Utah {failj, ABSTRACT Perceptual uer interface (PUI) are an important part of ubiquitou computing. Creating uch interface i difficult becaue of the image and ignal proceing knowledge required for creating claifier. We propoe an interactive machine-learning (IML) model that allow uer to train, claify/view and correct the claification. The concept and implementation detail of IML are dicued and contrated with claical machine learning model. Evaluation of two algorithm are alo preented. We alo briefly decribe Image Proceing with Crayon (Crayon), which i a tool for creating new camera-baed interface uing a imple painting metaphor. The Crayon tool embodie our notion of interactive machine learning. Categorie: H.5.2, D.2.2 General Term: Deign, Experimentation Keyword: Machine learning, perceptive uer interface, interaction, image proceing, claification INTRODUCTION Perceptual uer interface (PUI) are etablihing the need for machine learning in interactive etting. PUI like VideoPlace [8], Light Widget [3], and Light Table [15,16] all ue camera a their perceptive medium. Other ytem ue enor other than camera uch a depth canner and infrared enor [13,14,15]. All of thee PUI require machine learning and computer viion technique to create ome ort of a claifier. Thi claification component of the UI often demand great effort and expene. Becaue mot developer have little knowledge on how to implement recognition in their UI thi become problematic. Even thoe who do have thi knowledge would benefit if the claifier building expene were leened. We ugget the way to decreae thi expene i through the ue of a viual image claifier generator, which would allow developer to add intelligence to interface without forcing additional programming. Similar to how Viual Baic allow imple and fat development, thi tool would allow for fat integration of LEAVE BLANK THE LAST 2.5 cm (1 ) OF THE LEFT COLUMN ON THE FIRST PAGE FOR THE COPYRIGHT NOTICE. recognition or perception into a UI. Implementation of uch a tool, however, poe many problem. Firt and foremot i the problem of rapidly creating a atifactory claifier. The imple olution i to uing behind-thecene machine learning and image proceing. Machine learning allow automatic creation of claifier, however, the claical model are generally low to train, and not interactive. The claical machine-learning (CML) model i ummarized in Figure 1. Prior to the training of the claifier, feature need to be elected. Training i then performed off-line o that claification can be done quickly and efficiently. In thi model claification i optimized at the expene of longer training time. Generally, the claifier will run quickly o it can be done real-time. The aumption i that training will be performed only once and need not be interactive. Many machine-learning algorithm are very enitive to feature election and uffer greatly if there are very many feature. Feature Selection Train Interactive Ue Claify Figure 1 Claical machine learning model With CML, it i infeaible to create an interactive tool to create claifier. CML require the uer to chooe the feature and wait an extended amount of time for the algorithm to train. The election of feature i very problematic for mot interface deigner. If one i deigning an interactive technique involving laer pot tracking, mot deigner undertand that the pot i generally red. They are not prepared to deal with how to ort out thi pot from red clothing, camera noie or a variety of other problem. There are well-known image proceing feature for handling thee problem, but very few interface deigner would know how to carefully elect them in a way that the machine learning algorithm could handle. The current approach require too much technical knowledge on the part of the interface deigner. What we would like to do i replace the claical machine-learning

2 model with the interactive model hown in Figure 2. Thi interactive training allow the claifier to be coached along until the deired reult are met. In thi model the deigner i correcting and teaching the claifier and the claifier mut perform the appropriate feature election. Feature Selection Manual Correction Train Interactive Ue Claify Feedback To Deigner Figure 2 Interactive machine learning (IML) model The pre-election of feature can be eliminated and tranferred to the learning part of the IML if the learning algorithm ued perform feature election. Thi mean that a large repoitory of feature are initially calculated and fed to the learning algorithm o it can learn the bet feature for the claification problem at hand. The idea i to feed a very large number of feature into the claifier training and let the claifier do the filtering rather than the human. The human deigner then i focued on rapidly creating training data that will correct the error of the claifier. In claical machine learning, algorithm are evaluated on their inductive power. That i, how well the algorithm will perform on new data baed on the extrapolation made on the training data. Good inductive power require careful analyi and a great deal of computing time. Thi time i frequently exponential in the number of feature to be conidered. We believe that uing the IML model a imple viual tool can be deigned to build claifier quickly. We hypotheize that when uing the IML, having a very fat training algorithm i more important than trong induction. In place of careful analyi of many feature combination we provide much more human input to correct error a they appear. Thi allow the interactive cycle to be iterated quickly o it can be done more frequently. The remainder of the paper i a follow. The next ection briefly dicue the viual tool we created uing the IML model, called Image Proceing with Crayon (Crayon). Thi i done to how one application of the IML model power and veratility. Following the explanation of Crayon, we explore the detail of the IML model by examining it ditinction from CML, the problem it mut overcome, and it implementation detail. Finally we preent ome reult from ome tet between two of the implemented machine learning algorithm. From thee reult we bae ome preliminary concluion of IML a it relate to Crayon. IMAGE PROCESSING WITH CRAYONS Crayon i a ytem we created that ue IML to create image claifier. Crayon i intended to aid UI deigner who do not have detailed knowledge of image proceing and machine learning. It i alo intended to accelerate the effort of more knowledgeable programmer. There are two primary goal for the Crayon tool: 1) to allow the uer to create an image/pixel claifier quickly, and 2) to allow the uer to focu on the claification problem rather than image proceing or algorithm. Crayon i ucceful if it take minute rather than week or month to create an effective claifier. For implicity ake, we will refer to thi a the UI principle of fat and focued. Thi principle refer to enabling the deigner to quickly accomplih hi/her tak while remaining focued olely on that tak. Figure 3 how the Crayon deign proce. Image are input into the Crayon ytem, which can then export the generated claifier. It i aumed the uer ha already taken digital picture and aved them a file to import into the ytem, or that a camera i et up on the machine running Crayon, o it can capture image from it. Exporting the claifier i equally trivial, ince our implementation i written in Java. The claifier object i imply erialized and output to a file uing the tandard Java mechanim. Figure 3 Claifier Deign Proce An overview of the internal architecture of Crayon i hown in Figure 4. Crayon receive image upon which the uer doe ome manual claification, a claifier i created, then feedback i diplayed. The uer can then refine the claifier by adding more manual claification or, if the claifier i atifactory, the uer can export the claifier. The internal loop hown in Figure 4 directly correlate to the aforementioned train, feedback, correct cycle of the IML (ee Figure 2). To accomplih the fat and focued UI principle, thi loop mut be eay and quick to cycle through. To be interactive the training part of the loop mut take le than five econd and generally much fater. The cycle can be broken down into two component: the UI and the Claifier. The UI component need to be imple o the uer can remain focued on the claification problem at hand. The claifier creation need to be fat and efficient o the uer get feedback a quickly a poible, o they are not ditracted from the claification problem.

3 Figure 4 The claification deign loop Although the IML and the machine-learning component of Crayon are the primary dicuion of thi paper it i notable to mention that Crayon ha profited from work done by Viola and Jone [19] and Jaime and Chang [5,6,7]. Alo a brief example of how Crayon can be ued i illutrative. The equence of image in Figure 5 how the proce of creating a claifier uing Crayon. Figure 5 Crayon interaction proce Figure 5 illutrate how the uer initially paint very little data, view the feedback provided by the reulting claifier, correct by painting additional cla pixel and then iterate through the cycle. A een in the firt image pair in Figure 5, only a little data can generate a claifier that roughly learn kin and background. The claifier, however, over-generalize in favor of background; therefore, in the econd image pair you can ee kin ha been painted where the claifier previouly did poorly at claifying kin. The reulting claifier hown on the right of the econd image pair how the new claifier claifying mot of the kin on the hand, but alo claifying ome of the background a kin. The claifier i corrected again, and the reulting claifier i hown a the third image pair in the equence. Thu, in only a few iteration, a kin claifier i created. The implicity of the example above how the power that Crayon ha due to the effectivene of the IML model. The key iue in the creation of uch a tool lie in quickly generating effective claifier o the interactive deign loop can be utilized. MACHINE LEARNING For the IML model to function, the claifier mut be generated quickly and be able to generalize well. A uch we will firt dicu the ditinction between IML and CML, followed by the problem IML mut overcome becaue of it interactive etting, and latly it implementation detail including pecific algorithm. CML v. IML Claical machine learning generally ha the following aumption. There are relatively few carefully choen feature, There i limited training data, The claifier mut amplify that limited training data into excellent performance on new training data, Time to train the claifier i relatively unimportant a long a it doe not take too many day. None of thee aumption hold in our interactive ituation. Our UI deigner have no idea what feature will be appropriate. In fact, we are trying to inulate them from knowing uch thing. In our current Crayon prototype there are more than 150 feature per pixel. To reach the breadth of application that we deire for Crayon we project over 1,000 feature will be neceary. The additional feature will handle texture, hape and motion over time. For any given problem omewhere between three and fifteen of thoe feature will actually be ued, but the claifier algorithm mut automatically make thi election. The claifier we chooe mut therefore be able to accommodate uch a large number of feature, and/or elect only the bet feature. In Crayon, when a deigner begin to paint clae on an image a very large number of training example i quickly generated. With 77K pixel per image and 20 image one can rapidly generate over a million training example. In practice, the number tay in the 100K example range becaue deigner only paint pixel that they need to correct rather than all pixel in the image. What thi mean, however, i that deigner can generate a huge amount of training data very quickly. CML generally focue on the ability of a claifier to predict correct

4 behavior on new data. In IML, however, if the claifier prediction for new data are wrong, the deigner can rapidly make thoe correction. By rapid feedback and correction the claifier i quickly (in a matter of minute) focued onto the deired behavior. The goal of the claifier i not to predict the deigner intent into new ituation but rapidly reflect intent a expreed in concrete example. Becaue additional training example can be added o readily, IML bia differ greatly from that of CML. Becaue it extrapolate a little data to create a claifier that will be frequently ued in the future, CML i very concerned about overfit. Overfit i where the trained claifier adhere too cloely to the training data rather than deducing general principle. Cro-validation and other meaure are generally taken to minimize overfit. Thee meaure add ubtantially to the training time for CML algorithm. IML bia i to include the human in the loop by facilitating rapid correction of mitake. Overfit can eaily occur, but it i alo readily perceived by the deigner and intantly corrected by the addition of new training data in exactly the area that are mot problematic. Thi i hown clearly in Figure 5 where a deigner rapidly provide new data in the edge of the hand where the generalization failed. Our interactive claification loop require that the claifier training be very fat. To be effective, the claifier mut be generated from the training example in under five econd. If the claifier take minute or hour, the proce of train-feedback-correct i no longer interactive, and much le effective a a deign tool. Training on 100,000 example with 150 feature each in le than five econd i a eriou challenge for mot CML algorithm. Latly, for thi tool to be viable the final claifier will need to be able to claify 320 x 240 image in le than a fourth of a econd. If the reulting claifier i much lower than thi it become impoible to ue it to track interactive behavior in a meaningful way. IML Implementation Throughout our dicuion thu far, many requirement for the machine-learning algorithm in IML have been made. The machine-learning algorithm mut: learn/train very quickly, accommodate 100 to 1000 of feature, perform feature election, allow for ten to hundred of thouand of training example. Thee requirement put firm bound on what kind of a learning algorithm can be ued in IML. They invoke the fundamental quetion of which machine-learning algorithm fit all of thee criteria. We dicu everal option and the reaon why they are not viable before we ettle on our algorithm of choice: deciion tree (DT). Neural Network [12] are a powerful and often ued machine-learning algorithm. They can provably approximate any function in two layer. Their trength lie in their abilitie to intelligently integrate a variety of feature. Neural network alo produce relatively mall and efficient claifier, however, there are not feaible in IML. The number of feature ued in ytem like Crayon along with the number of hidden node required to produce the kind of claification that are neceary completely overpower thi algorithm. Even more debilitating i the training time for neural network. The time thi algorithm take to converge i far to long for interactive ue. For 150 feature thi can take hour or day. The nearet-neighbor algorithm [1] i eay to train but not very effective. Beide not being able to dicriminate amongt feature, nearet-neighbor ha eriou problem in high dimenional feature pace of the kind needed in IML and Crayon. Nearet-neighbor generally ha a claification time that i linear in the number of training example which alo make it unacceptably low. There are yet other algorithm uch a booting that do well with feature election, which i a deirable characteritic. While booting ha hown itelf to be very effective on tak uch a face tracing [18], it lengthy training time i prohibitive for interactive ue in Crayon. There are many more machine-learning algorithm, however, thi dicuion i ufficient to preface to our deciion of the ue of deciion tree. All the algorithm dicued above uffer from the cure of dimenionality. When many feature are ued (100 to 1000), their creation and execution time dramatically increae. In addition, the number of training example required to adequately cover uch high dimenional feature pace would far exceed what deigner can produce. With jut one deciion per feature the ize of the example et mut approach 2 100, which i completely unacceptable. We need a claifier that rapidly dicard feature and focue on the 1-10 feature that characterize a particular problem. Deciion tree [10] have many appealing propertie that coincide with the requirement of IML. Firt and foremot i that the DT algorithm i fundamentally a proce of feature election. The algorithm operate by examining each feature and electing a deciion point for dividing the range of that feature. It then compute the impurity of the reult of dividing the training example at that deciion point. One can think of impurity a meauring the amount of confuion in a given et. A et of example that all belong to one cla would be pure (zero impurity). There are a variety of poible impurity meaure [2]. The feature whoe partition yield the leat impurity i the one choen, the et i divided and the algorithm applied recurively to the divided ubet. Feature that do not provide dicrimination between clae are quickly

5 dicarded. The implicity of DT alo provide many implementation advantage in term of peed and pace of the reulting claifier. Quinlan original DT algorithm [10] worked only on feature that were dicrete (a mall number of choice). Our image feature do not have that property. Mot of our feature are continuou real value. Many extenion of the original DT algorithm, ID3, have been made to allow ue of real valued data [4,11]. All of thee algorithm either dicretize the data or by electing a threhold T for a given feature F divide the training example into two et where F<T and F>=T. The trick i for each feature to elect a value T that give the lowet impurity (bet claification improvement). The election of T from a large number of feature and a large number of training example i very low to do correctly. We have implemented two algorithm, which employ different diviion technique. Thee two algorithm alo repreent the two approache of longer training time with better generalization v. horter training time with poorer generalization. The firt trategy lightly reduce interactivity and relie more on learning performance. The econd relie on peed and interactivity. The two trategie are Center Weighted (CW) and Mean Split (MS). Our firt DT attempt wa to order all of the training example for each feature and tep through all of the example calculating the impurity a if the diviion wa between each of the example. Thi yielded a minimum impurity plit, however, thi generally provided a bet plit cloe to the beginning or end of the lit of example, till leaving a large number of example in one of the diviion. Diviion of thi nature yield deeper and more unbalanced tree, which correlate to lower claification time. To improve thi algorithm, we developed Center Weighted (CW), which doe the ame a above, except that it more heavily weight central plit (more equal diviion). By inuring that the plit threhold i generally in the middle of the feature range, the reulting tree tend to be more balanced and the ize of the training et to be examined at each level of the tree drop exponentially. CW DT do, however, uffer from an initial ort of all training example for each feature, reulting in a O(f * N log N) cot up front, where f i the number of feature and N the number of training example. Since in IML, we aume that both f and N are large, thi can be extremely cotly. Becaue of the extreme initial cot of orting all N training example f time, we have extended Center Weighted with CWSS. The SS tand for ub-ampled. Since the iteration through training example i purely to find a good plit, we can ample the example to find a tatitically ound plit. For example, ay N i 100,000, if we ample 1,000 of the original N, ort thoe and calculate the bet plit then our initial ort i 100 time fater. It i obviou that a better threhold could be computed uing all of the training data, but thi i mitigated by the fact that thoe data item will till be conidered in lower level of the tree. When a plit deciion i made, all of the training example are plit, not jut the ub-ample. The ub-ampling mean that each node plit deciion i never greater than O(f*1000*5), but that eventually all training data will be conidered. Quinlan ued a ampling technique called windowing. Windowing initially ued a mall ample of training example and increaed the number of training example ued to create the DT, until all of the original example were claified correctly [11]. Our technique, although imilar, differ in that the number of ample i fixed. At each node in the DT a new ample of fixed ize i drawn, allowing miclaified example in a higher level of the DT to be conidered at a lower level. The ue of ub-ampling in CWSS produced very light difference in claification accuracy a compared to CW, but reduced training time by a factor of at leat two (for training et with N 5,000). Thi factor however will continue to grow a N increae. (For N = 40,000 CWSS i approximately 5 time fater than CW; 8 for N = 80,000.) The CW and CWSS algorithm pend coniderable computing reource in trying to chooe a threhold value for each feature. The Mean Split (MS) algorithm pend very little time on uch deciion and relie on large amount of training data to correct deciion at lower level of the tree. The MS algorithm ue T=mean(F) a the threhold for dividing each feature F and compare the impuritie of the diviion of all feature. Thi i very efficient and produce relatively hallow deciion tree by generally dividing the training et in half at each deciion point. Mean plit, however, doe not enure that the diviion will necearily divide the example at point that are meaningful to correct claification. Succeive plit at lower level of the tree will eventually correctly claify the training data, but may not generalize a well. The reulting MS deciion tree are not a good a thoe produced by more careful mean uch a CW or CWSS. However, we hypotheized, that the peedup in claification would improve interactivity and thu reduce the time for deigner to train a claifier. We believe deigner make up for the lower quality of the deciion tree with the ability to correct more rapidly. The key i in optimizing deigner judgment rather than claifier prediction. MSSS i a ub-ampled verion of MS in the ame manner a CWSS. In MSSS, ince we jut evaluate the impurity at the mean, and ince the mean i a imple tatitical value, the reulting diviion are generally identical to thoe of traight MS. A a parenthetical note, another important bottleneck that i common to all of the claifier i the neceity to calculate all feature initially to create the claifier. We made the aumption in IML that all feature are pre-calculated and that the learning part will find the ditinguihing feature.

6 Although, thi can be optimized o it i fater, all algorithm will uffer from thi bottleneck. There are many difference between the performance of each of the algorithm. The mot important i that the CW algorithm train lower than the MS algorithm, but tend to create better claifier. Other difference are of note though. For example, the ub ampled verion, CWSS and MSSS, generally allowed the claifier to be generated fater. More pecifically, CWSS wa uually twice a fat a CW, a wa MSSS compared to MS. Becaue of the gain in peed and lack of lo of claification power, only CWSS and MSSS will be ued for comparion. The critical comparion i to ee which algorithm allow the uer to create a atifactory claifier the fatet. Uer tet comparing thee algorithm are outlined and preented in the next ection. EVALUATIONS Uer tet were conducted to evaluate the difference between CWSS and MSSS. When creating a new perceptual interface it i not claification time that i the real iue. The important iue i deigner time. A tated before, claification creation time for CWSS i longer than MSSS, but the center-weighted algorithm tend to generalize better than the mean plit algorithm. The CWSS generally take 1-10 econd to train on training et of 10,000-60,000 example, while MSSS i approximately twice a fat on the ame training et. Thee difference are important; a our hypothei wa that fater claifier creation time can overcome poorer inductive trength and thu reduce overall deigner time. To tet the difference between CWSS and MSSS we ued three key meaurement: wall clock time to create the claifier, number of claify/correct iteration, and tructure of the reulting tree (depth and number of node). The latter of thee three correpond to the amount of time the claifier take to claify an image in actual uage. In order to tet the amount of time a deigner take to create a good claifier, we need a tandard to define good claifier. A gold tandard wa created for four different claification problem: kin-detection, paper card tracking, robot car tracking and laer tracking. Thee gold tandard were created by carefully claifying pixel until, in human judgment, the bet poible claification wa being performed on the tet image for each problem. The reulting claifier wa then aved a a tandard. Ten total tet ubject were ued and divided into two group. The firt five did each tak uing the CWSS followed by the MSSS and the remaining five MSSS followed by CWSS. The uer were given each of the problem in turn and aked to build a claifier. Each time the ubject requeted a claifier to be built that claifier performance wa meaured againt the performance of the tandard claifier for that tak. When the ubject claifier agreed with the tandard on more than 97.5% of the pixel, the tet wa declared complete. Table 1, how the average time and iteration for the firt group, Table 2, the econd group. CWSS Problem Time Iteration MSSS Time Iteration Skin 03: : Paper Card 02: : Robot Car 00: : Laer 00: : Table 1 CWSS followed by MSSS MSSS Problem Time Iteration CWSS Time Iteration Skin 10: : Paper Card 04: : Robot Car 01: : Laer 01: : Table 2 MSSS followed by CWSS The laer tracker i a relatively imple claifier becaue of the uniquene of bright red pot [9]. The robot car wa contrated with a uniform colored carpet and wa imilarly traightforward. Identifying colored paper card againt a cluttered background wa more difficult becaue of the diverity of the background. The kin tracker i the hardet becaue of the diverity of kin color, camera overaturation problem and cluttered background [20]. A can be een in table 1 and 2, MSSS take ubtantially more deigner effort on the hard problem than CWSS. All ubject pecifically tated that CWSS wa fater than MSSS epecially in the Skin cae. (Some did not notice a difference between the two algorithm while working on the other problem.) We did not tet any of the lower algorithm uch a neural net or nearet-neighbor. Interactively thee are o poor that the reult are elfevident. We alo did not tet the full CW algorithm. It claification time tend into minute and clearly could not compete with the time hown in table 1 and 2. It i clear from our evaluation that a claification algorithm mut get under the econd barrier in producing a new claification, but that once under that barrier, the deigner time begin to dominate. Once the deigner time begin to dominate the total time, then the claifier with better generalization win.

7 We alo mentioned the importance of the tree tructure a it relate to the claification time of an image. Table 3 how the average tree tructure (tree depth and number of node) a well a the average claification time (ACT) in milliecond over the et of tet image. Problem Dept h CWSS Node ACT Dept h MSSS Node ACT Skin Paper Card Car Laer Table 3 Tree tructure and average claify time (ACT) A een in Table 3, depth, number of node and ACT, were all lower in CWSS than in MSSS. Thi wa predicted a CWSS provide better diviion between the training example. While teting we oberved that thoe who ued the MSSS which i fat but le accurate, firt, ended up uing more training data, even when they ued the CWSS, which uually generalize better and need le data. Thoe who ued the CWSS firt, were pleaed with the interactivity of CWSS and became very frutrated when they ued MSSS, even though it could cycle fater through the interactive loop. In actuality, becaue of the poor generalization of the mean plit algorithm, even though the claifier generation time for MSSS wa quicker than CWSS, the uer felt it neceary to paint more uing the MSSS, o the overall time increaed uing MSSS. CONCLUSION When uing machine learning in an interactive deign etting, feature election mut be automatic rather than manual and claifier training-time mut be relatively fat. Deciion Tree uing a ub-ampling technique to improve training time are very effective for both of thee purpoe. Once interactive peed are achieved, however, the quality of the claifier generalization become important. Uing tool like Crayon, demontrate that machine learning can form an appropriate bai for the deign tool needed to create new perceptual uer interface. REFERENCES 1. Cover, T., and Hart, P. Nearet Neighbor Pattern Claification. IEEE Tranaction on Information Theory, 13, (1967) Duda, R. O., Hart, P. E., and Stork, D. G., Pattern Claification. (2001). 3. Fail, J.A., Olen, D.R. LightWidget: Interacting in Everyday Space. Proceeding of IUI 02 (San Francico CA, January 2002). 4. Fayyad, U.M. and Irani, K. B. On the Handling of Continuou-valued Attribute in Deciion Tree Generation. Machine Learning, 8, ,(1992). 5. Jaime, A. and Chang, S.-F. A Conceptual Framework for Indexing Viual Information at Multiple Level. IS&T/SPIE Internet Imaging 2000, (San Joe CA, January 2000). 6. Jaime, A. and Chang, S.-F. Automatic Selection of Viual Feature and Claifier. Storage and Retrieval for Image and Video Databae VIII, IS&T/SPIE (San Joe CA, January 2000). 7. Jaime, A. and Chang, S.-F. Integrating Multiple Claifier in Viual Object Detector Learned from Uer Input. Invited paper, eion on Image and Video Databae, 4th Aian Conference on Computer Viion (ACCV 2000), Taipei, Taiwan, January 8-11, Krueger, M. W., Gionfriddo. T., and Hinrichen, K., VIDEOPLACE -- an artificial reality. Human Factor in Computing Sytem, CHI '85 Conference Proceeding, ACM Pre, 1985, Olen, D.R., Nielen, T. Laer Pointer Interaction. Proceeding of CHI 01 (Seattle WA, March 2001). 10. Quinlan, J. R. Induction of Deciion Tree. Machine Learning, 1(1); , (1986). 11. Quinlan, J. R. C4.5: Program for machine learning. Morgan Kaufmann, San Mateo, CA, Rumelhart, D., Widrow, B., and Lehr, M. The Baic Idea in Neural Network. Communication of the ACM, 37(3), (1994), pp Schmidt, A. Implicit Human Computer Interaction Through Context. Peronal Technologie, Vol 4(2), June Starner, T., Auxier, J. and Ahbrook, D. The Geture Pendant: A Self-illuminating, Wearable, Infrared Computer Viion Sytem for Home Automation Control and Medical Monitoring. International Sympoium on Wearable Computing (Atlanta GA, October 2000). 15. Trigg, B. Model-baed Sonar Localiation for Mobile Robot. Intelligent Robotic Sytem 93, Zakopane, Poland, Underkoffler, J. and Ihii H. Illuminating Light: An Optical Deign Tool with a Luminou-Tangible Interface. Proceeding of CHI 98 (Lo Angele CA, April 1998). 17. Underkoffler, J., Ullmer, B. and Ihii, H. Emancipated Pixel: Real-World Graphic in the Luminou Room. Proceeding of SIGGRAPH 99 (Lo Angele CA, 1999), ACM Pre, Vailaya, A., Zhong, Y., and Jain, A. K. A hierarchical ytem for efficient image retrieval. In Proc. Int. Conf. on Patt. Recog. (Augut 1996).

8 19. Viola, P. and Jone, M. Robut real-time object detection. Technical Report 2001/01, Compaq CRL, February Yang, M.H. and Ahuja, N. Gauian Mixture Model for Human Skin Color and It Application in Image and Video Databae. Proceeding of SPIE 99 (San Joe CA, Jan 1999),