Session B11 #122 Disclaimer—This paper partially fulfills a writing requirement for first year (freshman) engineering students at the University of Pittsburgh Swanson School of Engineering. This paper is a student, not a professional, paper. This paper is based on publicly available information and may not provide complete analyses of all relevant data. If this paper is used for any purpose other than these authors’ partial fulfillment of a writing requirement for first year (freshman) engineering students at the University of Pittsburgh Swanson School of Engineering, the user does so at his or her own risk. THE APPLICATION OF MACHINE LEARNING IN FACIAL RECOGNITION Keting Zhao, [email protected], Mahboobin 4:00, Jiacong Liu, [email protected], Mahboobin 4:00 Abstract— Most Facial Recognition(FR) systems are using principal component analysis which limits the accuracy and efficiency of the recognition process. Thus, a better algorithm is needed to improve the current FR system. This paper introduces the Machine Learning based FR system and discusses its potential social impacts. Integrating the Machine Learning algorithm into the FR system will allow the computer to learn and detect more detailed differences among given objects. In that, the optimized FR system will yield better image search results with a lower error rate. This application of Machine Learning is significant to strengthen the law enforcement and to improve the surveillance system as well. However, some ethical issues also arise with this optimized FR system, mainly on violating the individual privacy in public. These ethical issues can influence on the sustainability of FR system. Furthermore, the solution of these moral issues is the key to improve its sustainability. Key Words— Computing ethics, Computer Object Recognition, Facial Recognition, Machine Learning Algorithm, Public security and privacy INTRODUCTION:THE CURRENT EXISTING ISSUE IN FACIAL RECOGNITION Facial recognition(FR) uses the computer to determine individuals by their facial features and has been applied in various field in the past decade. As indicating in the TV show, Person of Interest, FR system could predict the criminal actions and sends out the corresponding information of the targeted person such as his or her social security number. The algorithm that facial recognition used in the past was called Principle Component Analysis, which simply transformed various human facial features, such as the distance between the eyes and nose, into geometric shapes. The software would then match these shapes with those on actual faces in order to recognize an individual’s identity. The problem with this analysis, however, was it limited facial recognition accuracy since it could not distinguish between two people with similar facial patterns like human beings can. Therefore, a better algorithm was University of Pittsburgh Swanson School of Engineering 1 03.03.2017 needed. The technology modified to solve this issue is called the Machine Learning Algorithm. In the next section, the basic mechanism of Machine Learning will be introduced. Further in the paper, FR system and it relevant application and case study will be showed as well as the current ethical issues that FR is facing. MECHANISMS OF MACHINE LEARNING For a human, “learning” means to gain knowledge or skills through study, experience, or being taught. Therefore, the general idea of machine learning algorithms is to let the computer “understand” objects, such as a human face, through the learning process similar to the human learning process. Two important reason for a machine to have the ability of learning is that: concise the relationship between input and output among large amount of data, which could not be done by human; extract the significant relationships among the hidden piles of data [1]. In other words, machine learning method could detect the desired patterns from the providing enormous data sets and classify, or separate the objects into specific classes, or category. Data Structure and Machine Learning There are several parallels between machine learning computational and human models are developed in the way that is based on the theory of animal and human learning [2]. The basic process in machine learning is just like how humans learn the structure of the universe [3]. Let’s say a person lives in a twodimensional random flat world where every pixel value, or every point that builds up this flat world, is random across the time and space. It is hard for the person to point out which direction is north or south because there is nothing makes any part of this universal distinguishable from other parts. Therefore, the idea of moving north or south in this two-dimensional world is meaningless since the knowledge of north and south does not exist in this world. Thus, before a human or the machine to learn the structure of the universe, there has to have the structure in the universe which is correlated with some patterns and is relational within time and space between these pixel Keting Zhao Jiacong Liu value in order to have the knowledge that could be learned. Referring this back into the example of moving north and south is that the different magnetic pole given to north and south is their unique structure which human learned and use to define north and south eventually. In other words, the fundamental of the machine learning algorithm is to learn the structure in the data; when specific structure does exist in the data, the machine which works as human brain will then able to catch the pattern in the structure, and determine the object based on their special correlations or compressing the data structure into corresponding geometric shapes based on their arrangement. Through combining a lot of correlated information across different modalities, such as fitting these pattern into social construct and emotions, a human starts to “create” reality and derive his or her own selfconsciousness [3]. The computer using a Machine Learning algorithm in the similar way to learn the structure of data and analysis the pattern in the data structure and eventually derive appropriate result under giving contexts. In a generally terms, whenever machine changes its structure, database(inputs), and program in such a manner that improves its performance on such tasks, it is considered machine learning. The figure 1 shows the basic components that the Machine Learning Algorithm needs to include and any of the improvement occurs in these components count as learning. dimension represents one distinguishable information of the targeted object in the real world. By using vectors to sort attributes, or information, each object could be represented by a point in an enormous large database. For example, in order to represent individual cats and dogs at a pet store in a two-dimensional data base, the data analysist will set y-axis as tail length and x-axis as its weight. Thus, on a x-y coordinates graph, each dog and cat will be store as a point with the corresponding (x, y) based on its weight and tail length. When more attributes of these cats and dogs are added in, they are still presented as points just in a high coordinate, such as (x, y, z, h, j). In general, the first step in the machine learning is to reduce all the features of the objects into data points as the appropriate input for the machine learning algorithm. Supervised Learning in Machine Learning Algorithm One of the task that machine learning algorithm tries to accomplish in the context above is to distinguish which points are cats and which points are dogs. As showing in figure 2, red points are cats and blue points are dogs. With human visual processing system, a person could easily learn a dog or a cat by looking its multiple feature or the data structure. However, the computer does not have this such advanced visual system. Thus, the machine learning algorithm is going to draw the classification function or the separation line based on the labels of data points, in this case, the green line in the graph. The computer takes these input data which has labels described as a cat or a dog and generates the relationship among these data points which is the way of computer learning. In that, when a new vector sorts into this data base, the computer could use the relationship it generates earlier to be able to accurately determine if this vector represents a dog or a cat. FIGURE 1[1] The diagram of AI System Using of Vectors in the Machine Learning FIGURE 2[15] Plotting graph of cats and dogs at the pet store (Red spots - Cat, Blue spots - Dogs) The fundamental mathematics in the machine learning algorithm is the vector which is the real world in the eyes of the computer [1]. A vector is another way to say a point in multiple dimensional space where each 2 Keting Zhao Jiacong Liu The example above describe a branch of machine learning, called supervised learning. By definition, supervised learning means to generate a function based upon assigned labels that maps input to desired output [1]. This means the initial training using data set has been given classification already and the relationship is generated from the data set could be used as a useful tool to determine the class of new input data. This is learning algorithm has been used in facial recognition system which will be discussed in later section. Unsupervised Learning in Machine Learning Algorithm FIGURE 4[12] Gaussian mixture model Example Another type of learning mechanism in machine learning algorithm will be also applied in facial recognition system called unsupervised learning [1], which applies when the given data does not have label on them. In this case, the computer needs to determine whether there is specific data structure in the data sets or not. The particular functions using in the Machine Learning Algorithm are K-means clustering and Gaussian mixture model [3]. Figure 3 and Figure 4 are examples of using these two function in the Machine Learning. Kmeans clustering function helps to find the relational and correlational associations and distances between the vector values. The computer will come out the result that when the distances between two data points are lower than certain threshold, the relationship determines as close and if not, then the relationship will be determines as far. Thus, all the data points will be separate into different cluster based on the structure which just native to the data. Similarly, the data could be separate into groups based on their histogram value and compute into Gaussian mixture model. Thus with unsupervised learning, the computer could generate the relationship among data independently to the labeling. With this mechanism, the FR system will gain more flexibility and eventually process the recognition progress as a human and even faster and more accurately than a human. FACIAL RECOGNITION WITH PCA (PRINCIPAL COMPONENT ANALYSIS) In the past decade scientists have made several different approaches to build Facial Recognition systems through combining complex mathematical algorithms with Machine Learning. Among them one of the most popular solutions was the Facial Recognition system based on the Principal component analysis (PCA). The PCA was an algorithm that was designed to train computers to recognize similar patterns between objects (in this case objects would be images). This algorithm used a mathematical formula to lower the dimension of images and stored the information of these images into vectors. Doing so allowed computers to manipulate the image information for further analysis. The PCA recognition system was widely adapted and used by agencies such as FBI for criminal verification and used as employee check-in system for ordinary companies. Using criminal verification as an example, suppose a police station has 1000 face images of 100 criminals collected from past criminal records, and the police wanted to use these images as a criminal database to verify whether a person has committed crimes in the past. To present such a task, the operator must manually input this dataset full with each faces labeled into the PCA recognition system, and split the images into different groups. For this scenario, suppose the operator split the data into 100 groups, each groups with 10 images from the same identity into the PCA system. FIGURE 3[12] K-means Clustering Example 3 Keting Zhao Jiacong Liu FIGURE 7 [4] Example of Eigenfaces Once the eigenvector is transformed into a normal size image, the operator would get a so-called “Eigenface.” The top four pictures in figure 7 are examples of eigenfaces. In this case the police station would have 100 eigenfaces stored in the database and ready to be used. FIGURE 5 [4] PCA Image processing 1 FIGURE 8 [4] Determining Similarity Through Linear Algebra When the operator inputs a new face image into the system, the system would again transform the new image into vector, then compare the vector to each of the eigenvector and compute the similarity between the new face and the existing eigenvectors as shown in figure 8 [4]. If the computer found a high similarity between input vector and one of the eigenvectors, it will look further into the group related to that eigenvector and find out which identity specifically matched with the input vector. On the other hand, if the computer found the input vector has no similarity with all of the eigenvectors in the database, the system would conclude that this person has not been recorded for crime in the past. FIGURE 6 [4] PCA Image processing 2 As figure 5 and, 6 show above, the PCA algorithm takes all input images and transform them into a 2D vector [4]. Each value inside the vector correlates to the pixel in the original image. Then, it would combine all the 2D vectors that belong to the same group into a single vector to represent all identities from the group. By taking the average value and normalizing the group vectors, the system will create an eigenvector for each group [4]. Eigenvector could be understood as a vector that contains value of the “average” face of all 10 different faces from one group. Sub- Disadvantage, Shortcoming of PCA-like System Although statistics have shown that the PCA system has an accuracy rate of 60%-80% [6], there are many shortcomings and conditions that limit the performance of the PCA algorithm. First, the lightning condition of the image have to be perfect. In order for computers to conduct comparisons at maximum level, the faces in the image must be fully under good lighting conditions thus all the pixels can be transformed into useful values in the vector. In poor light conditions only part of the face would be revealed which would heavily impact the success rate in matching. Secondly, the PCA algorithm required all the pictures in the database and the input images to be the exact same size with 4 Keting Zhao Jiacong Liu the face centered at a fixed position. The requirement in alignment restricted the flexibility of the recognition system, forcing the operator to manually crop and adjust all the pictures that didn’t fit the standard. Last, all pictures must be frontal image of faces, which means if the face in the image were tilted at an angle the system might failed the matching test [4]. Given these pre-existing flaws in the PCA and PCAlike systems, a better facial recognition algorithm is in demand to improve the accuracy and usability of current machine recognition systems. capture six reference points on the detected face, then it crops the image to a suitable size keeping only the general face image to analyze (9b). To compute a 3D model based on a 2D image, computers will then find 67 additional reference points on the cropped image (9c), projecting the facial features into a 3D model (9d, 9e). Doing so not only allow computers to rotate the face in the image to a proper angle for recognition (9g) but could also be used to predict the side look of the face (9h). This 3D modeling method solved the facial alignment problem in the past facial recognition systems and it increased the usability of the program since now the system can recognize someone from a picture even if the face is not fully frontal [5]. FACIAL RECOGNITION BASED ON DEEP MACHINE LEARNING (DEEPFACE) In 2014, Facebook announced their own facial recognition system – DeepFace, claiming that it has a 97% accuracy in recognizing human faces. In a paper published by the Facebook AI Research lab, it stated that the Facebook DeepFace Algorithm is different from other facial recognition systems because it used the 3D face alignment and Deep Machine Learning algorithm. FIGURE 10 [5] Example of Nine-Layer DNN What Is DeepFace Unlike the PCA algorithm, where an operator has to crop and edit all pictures following a certain guideline, the DeepFace algorithm trained computers to detect the coordinates of the face in a picture by themselves. Being a social network company itself, Facebook has one of the largest image database in the world, which enabled the DeepFace team show computers millions of face images and let computers understand and learn the overall structure of a human face through mathine learning algorithm [5]. This allowed computers to detect faces and create a frontal image of the face from undocumented pictures all by itself. In terms of processing the face image, instead of transforming the entire image into vector like the PCA system would, DeepFace uses a nine-layer neural network to extract and process useful information of the image (figure 10). Eventually the system will compute a vector that represents the input face image (figure 10, F7), but every value in the vector would only be a useful representation of the facial features such as distance between eyes and nose [5]. This comparison method is much more reliable than PCA algorithm since the representation vector created by DeepFace doesn’t contain any useless data. DeepFace also built its database based on labeled faces. First it grouped all images with the same labeled identity together and studied their similarities. Then the system computes and stores the representation vector for the face of that identity for future references. When the operator inputs an undocumented image asking the system to verify the identity, the system would compare the vector of the new input image with existing representation vectors in the database and return the identity once a matched is found or conclude the identity is unknown [5]. In this sense the DeepFace would has a much higher accuracy rates than the PCA system since DeepFace really understands the structure of a human face and only takes the important parameters into account for comparison. According to Labeled Faces in the Wild (LFW) benchmark, an online image database specifically designed to test facial recognition systems, the DeepFace algorithm achieved an accuracy rate of 97.35%, while the PCA algorithm only scored with a 60.2% [6]. The DeepFace algorithm have enabled computers to recognize faces at the human level. FIGURE 9 [5] DeepFace Analyzing Image Figure 9a is a simulation of what DeepFace system “sees” from an image. In DeepFace recognition system, computers will find the general area of the human face and 5 Keting Zhao Jiacong Liu APPLICATION OF FACIAL RECOGNITION Surveillance There are many fields of interest in our society that facial recognition can be apply to. The two main applications of the facial recognition include identification and surveillance. These two applications will benefit from the improved facial recognition system. Facial recognition system could also greatly improve the current surveillance system. Today there are hundreds of surveillance cameras everywhere, but none of them can actually tell the identity of a person. With that being said, all the surveillance cameras are useless without people seating behind the monitors and staring at all the faces that are passing by. However, with facial recognition system, law enforcement could use computers and surveillance cameras to locate suspects and criminals. In fact, back in 2001, the PCA facial recognition system was experimented in monitoring all the people entering the super bowl game and found 19 criminals [9]. Although due to the uncertainty of accuracy, movement and poses of the subjects, and public concerns over privacy violation, FBI and police departments have given up expanding the usage of facial recognition on surveillance. The PCA recognition system required the subject to sit still and fully face the camera in order the perceive a usable face image, but the problem here was that since people were moving all the time, the PCA system didn’t work all that well. This is where DeepFace algorithm can make up for the disadvantages of the PCA algorithm by applying its 3D modeling, even when the people are constantly moving or facing the camera sideways, it would still be able to capture the face image. The same principle as finding criminals, facial recognition can also be applied to finding missing persons. For example, if a child got lost in Disneyland, the parents of the child could provide a sample picture of the child and input them into the facial recognition system; then by using the surveillance camera, the system would locate the position of the child. With its high accuracy, DeepFace has many potential usages that could provide countless values to the society in the future. The law enforcements could use it to greatly improve our security. Identification The main objective of identification is to solve the general problems such as “Who is this?” and “Is this the right person?” Solving these questions is essential in scenarios where identification is used to gain permission for accessing personal properties. Facial recognition could be used similarly to fingerprints: allowing you to unlock your cell phone and set up payment method based on your identity. The following are some examples of these application in real life scenarios. First, imagine a door lock with pre-installed facial recognition system that would unlock the door when it recognized the owner of the house; you would no longer need to carry a key around and worryed about losing it. Second, experiments on facial payment have been conducted by a Finnish software company called Uniqul [7]. In the future, people can link their payment methods and bank accounts together with their unique faces. Once all the information has been recorded, one can go to a grocery store or a movie theater without bringing a wallet, and pay simply by looking at the camera for a few seconds. When the system recognizes the face, it will make the transaction online. We no longer will have to pull out our credit card, enter passwords, and sign the receipt when going to a store. Another potential application of the facial recognition system is to reinforce academic integrity. The college standardized tests such as ACT and SAT could use the facial recognition to verify valid test takers and prevent substitution test takers from entering the exam room. This would promise an equal chance for all test takers. Last, in China, several tourist attractions limit the number of people that can enter the site each day because some of these attractions are old and need to be preserved. Limiting the number of tourists would make it a lot easier to maintain the attraction sites. However, limiting the number of tickets provide scalpers an opportunity to raise the price and make profit from reselling these tickets. To solve this problem, WuZhen, one of the oldest tourist sites, decided to use facial recognition to verify the owner of the ticket, thus only the person who bought the ticket would have the permission to enter the tourist site [8]. These scenarios are only valid and applicable with the assumption that facial recognition system are accurate to a standard that it won’t make mistakes. Clearly the DeepFace would outperform PCA algorithm in these scenarios with a human level accuracy. ETHICAL CONCERNS AND SUSTAINABILITY OF MACHINE LEARNING BASED FACIAL RECOGNITION The current argument about the Machine Learning based Facial Recognition is the tradeoff among security, privacy and freedom [10]. The balance among these three elements will also determine the sustainability of FR system. United Nations Conference on Sustainable Development emphasizes that sustainability is not just strong economic performance but intergenerational and intergenerational equity which involves a balanced consideration of social, economic and environmental goals and objectives in both public and private decision-making [11]. In this case, how to reach the equity at the tradeoff of personal privacy, freedom and public security is the main roadblock of the 6 Keting Zhao Jiacong Liu sustainability of optimized FR system. Brey, from University of Twente, indicates in his journal that violation of freedom and violation of privacy are two main moral issues that have effect on the sustainability of FR system. Since the Machine Learning based Facial Recognition System is designed to connect with online databases, which contain sufficient amounts of personal information such as social security number, medical prescription, and ID pictures on driver license, the potential violation of personal privacy at public space becomes an issue. Although the opponents do not deny the concept that the optimized FR could reduce the crime rate and enhance the life quality, they question its reliability and efficiency of stopping crime [10]. Mainly, opponents argue that the trends of leaking personal information to the illegal users and error rates of identifying criminals are great than the gains in security. On the other hands, if engineers could find the solutions of these moral concerns, the sustainability of FR systems will be improved. Bedoya, executive director of the center on privacy and technology at Georgetown Law, claims that, “No federal law controls this technology, no court decision limits it. This technology is not under control” [a]. FR system will be used by different agencies for slightly different purpose. For example, FBI will use it to determine a criminal which needs the database that contains all the criminal records, while other police department will use it to trace a missing high school students which will need to access the database that contains information of high students. It is understandable that determining which databases are necessary for the tasks is hard. However, sometimes the law enforcement will use the databases without notifying and this put the sustainability of FR system on risk of violating individuals’ privacy. FBI had been revealed that they first launched its advanced biometric database, Next Generation Identification, in 2010, enlarging the previous fingerprint database with further capabilities including FR system. However, the bureau did not inform the public about its newfound capabilities nor did it publish a privacy impact assessment, required by law, for five years. Furthermore, unlike with the collection of fingerprints and DNA, which is done after the legal arrestments, photos of innocent civilians are being collected proactively. The FBI made arrangements with 18 different states to gain access to their databases of driver’s license photos [12]. It is appalling that one’s picture on his or her driver license will be put into a repository that could be searched by law enforcement across the country. The purpose to optimize FR system is to provide better security for the society, but the action of FBI violates this purpose and put the entire society under unnecessary panic that each personal information has been scanned and used without noticing. If this ethical issue cannot be solved, the negative feedback from the society will overwrite the benefits that FR could bring to the surveillance system. This unbalance will eventually damage the sustainability of the FR system. In addition, the growing flexibility in FR system itself will also cause violation of privacy. FR system has the ability of doing contextual integrity [10], relating an identified person with their other personal information by aggregating data from multiple database. All the confidentially agreements are invalid at this point. Since there is also no clear standard to determine who could be users of FR system [13], some data collectors may easily get access to the system and sell the aggregative information which originally under the protection of confidentially agreements. For example, information about people’s prescription may be sold to medicine companies for marketing purposes, or information collected for scientific purposes may be used in some political activities. In addition, there is also no rules about the purpose of using of the FR system which may cause domain and user shifts [10]. For instance, instead of using the FR system to examine criminal suspects, a police force could use it to do loop statistical analysis on a composition of crowd’s face prints in order to track individuals over long Violation of Freedom When the FR yields the incorrect matches and sends the wrong alert to the police, the payment would be that the innocent citizens become subjected to harassment by police [10]. The opponents claim that this action violates individual freedom; nobody deserves to experience arrestment for crimes he or she does not commit. People do tend to accept the idea that it has to suffer minor inconveniences so that criminals can be apprehended. However, under the current situation, the harm done to innocent citizens who are determined as false positive may begin to outweigh the benefits of a few additional arrests of criminals. Engineers reported that there are three types of error which are responsible for the occurrence of incorrect matches in optimized FR [10]. First error exists when unpredictable wrong information appears in the online data base where the FR abstracts the inputs from. Second, the errors in probability estimates creates a margin of error which is not avoidable. The third type of error occurs due to the wrong way of installing and using the system. In order to narrow down the false positive results, it requires to intergrade more accurate and comprehensive data check codes in the program. For example, we think to add error check points or multiple data comparison from different database in the algorithm before using the data as input to the FR system. If the false positive results could be significantly narrowed down, the value of FR system towards law enforcement can maintain its social sustainability which, in other words, it can actually enhance the public security by reporting the correct criminals to the police. Violation of Privacy Another debating on FR system is its assess to different databases. So far, there is no regulation clear states which databases FR system could have access to. As Alvaro 7 Keting Zhao Jiacong Liu distance. As journalist Richard Meares reports, there have been several reports of damaging of the FR system due to abuse the system repeatedly by tracking and zooming in on attractive women [10]. Opponents argue that this action will cause panic in public with the uncomfortable feeling of being watched incessantly. Thus, the violation of privacy could be limited by setting strict regulation of using as well as by setting user identification checkpoint before login the system. Adding obligations of preventing violation of privacy on system developers and users can also be the efficient way to stop the undesirable forms of violation occurring. To stop these will not only requires continuously optimizing the algorithm of FR system, but also requires the government putting efforts to establish the standard of using FR system. If a satisfied balance of the tradeoff between public security and individuals’ freedom and privacy could be reached, the FR system will have a sustainable development seven toward next generation. https://research.fb.com/wpcontent/uploads/2016/11/deepface-closing-the-gap-tohuman-level-performance-in-face-verification.pdf [6] “Labeled Faces in the Wild.” University of Massachusetts. n.d. Web. Accessed 26 Feb 2017 http://vis-www.cs.umass.edu/lfw/results.html [7] “World’s first face recognition payment method.” Uniqul, Inc. 15 Jul. 2013. Web. Accessed 28 Feb 2017. http://uniqul.com/worlds-first-face-recognition-paymentsystem/ [8] T. Revell. “Chinese tourist town uses face recognition as an entry pass.” New Scientist. Reed Business Information, Ltd. 26 Nov. 2016. Web. Accessed 1 Mar. 2017. https://www.newscientist.com/article/2113176-chinesetourist-town-uses-face-recognition-as-an-entry-pass/ [9] V. Chachere. “Biometrics Used to Detect Criminals at Super Bowl.” ABC News. 13 Feb. 2002. Web. Accessed. 2 March. 2017. http://abcnews.go.com/Technology/story?id=98871 [10] Brey, P. "Ethical Aspects of Facial Recognition Systems in Public Places." Journal of Information, Communication & Ethics in Society (2004): 97-109. University of Twente. University of Twente. Web. 26 Jan. 2017. https://www.utwente.nl/en/bms/wijsb/staff/brey/Publicati es_ Brey/Brey_2004_Face-Recognition.pdf [11]"Sustainable Development Goals: 17 Goals to Transform Our World." United Nations. United Nations. Web. 31 Mar. 2017. http://www.un.org/sustainabledevelopment/ [12] Solon, Olivia. "Facial recognition database used by FBI is out of control, House committee hears." The Guardian. Guardian News and Media, 27 Mar. 2017. Web. Accessed 31 Mar. 2017. https://www.theguardian.com/technology/2017/mar/27/u s-facial-recognition-database-fbi-drivers-licenses-passports [13] Padania, Sameer. "The Ethics of Face Recognition Technology." WITNESS Blog. N.p., 20 July 2015. Web. Accessed 11 Jan. 2017. https://blog.witness.org/2012/03/the-ethics-of-facerecognition-technology/ [14] Maciej Pacula. N.p., n.d. Web. 03 Mar. 2017. http://blog.mpacula.com/2011/04/27/k-means-clusteringexample-python/ PERCPECTION ON MACHINE LEARNING BASED FACIAL RECOGONITION Based on deep machine learning, the DeepFace FR system is the ideal solution to overcome the difficulties and flaws encountered by previous FR systems. If used correctly, the DeepFace system could provide numerous valuable application to strengthen the verification, security, and surveillance system in our society. However, there must be regulations and standards established to restrict the area of interest for using FR systems in order to prevent violation on individual privacies and freedoms. SOURCES [1] Nillson, Nils. “Introduction to Machine Learning.” Stanford University. N.p., 4 February 2015. Web. Accessed 10 February 2017. http://ai.stanford.edu/~nilsson/MLBOOK.pdf [2] Jordan, M., and T. Mitchell. "Machine learning: Trends, perspectives, and prospects." Science AAAS. N.p., 17 June 2015. Web. Accessed 12 Jan. 2017. http://science.sciencemag.org/content/349/6245/255.full [3] TheScienceguy3000. YouTube, 27 July 2013. Web. 03 Mar. 2017. https://www.youtube.com/watch?v=-rMMTv7XLYw [4] M. Turk and A. Pentland. “Eigenfaces for Face Detection/Recognition.” Journal of Cognitive Neuroscience. 1991. Accessed 25 Feb. 2017. [5] Y. Taigman, M. Yang, M. Ranzato, and L. Wolf. “DeepFace: Closing the Gap to Human-Level Performance in Face Verification” Facebook AI Research Lab. Facebook, Inc. 2016. Web. Accessed 25 Feb 2017. ACKNOWLEDGMENTS Thank you to our writing instructor, Rachel for continuously giving us useful feedback and pushing us to write a better paper. Also, thank you to our co-chair, Patrick, for regularly checking on our process stratus. Once more, thank you to Beth for talking us through all the information and requirements of this conference paper. 8
© Copyright 2026 Paperzz