{"title": "Who's In the Picture", "book": "Advances in Neural Information Processing Systems", "page_first": 137, "page_last": 144, "abstract": null, "full_text": "                            Who's in the Picture?\n\n\n\n          Tamara L. Berg, Alexander C. Berg, Jaety Edwards and D.A. Forsyth\n                                 Computer Science Division\n                                        U.C. Berkeley\n                                     Berkeley, CA 94720\n                                   millert@cs.berkeley.edu\n\n\n\n                                          Abstract\n\n          The context in which a name appears in a caption provides powerful cues\n          as to who is depicted in the associated image. We obtain 44,773 face im-\n          ages, using a face detector, from approximately half a million captioned\n          news images and automatically link names, obtained using a named en-\n          tity recognizer, with these faces. A simple clustering method can pro-\n          duce fair results. We improve these results significantly by combining\n          the clustering process with a model of the probability that an individual\n          is depicted given its context. Once the labeling procedure is over, we\n          have an accurately labeled set of faces, an appearance model for each\n          individual depicted, and a natural language model that can produce ac-\n          curate results on captions in isolation.\n\n\n\n1    Introduction\n\nIt is a remarkable fact that pictures and their associated annotations are complementary.\nThis observation has been used to browse museum collections ([1]) and organize large\nimage collections ([2, 7, 12, 13]). All of these papers use fairly crude \"bag of words\"\nmodels, treating words as floating tags and looking at the co-occurrence of image regions\nand annotated words. In this paper, we show that significant gains are available by treating\nlanguage more carefully.\n\nOur domain is a large dataset of news photographs with associated captions. A face detector\nis used to identify potential faces and a named entity recognizer to identify potential names.\nMultiple faces and names from one image-caption pair are quite common. The problem is\nto find a correspondence between some of the faces and names. As part of the solution\nwe learn an appearance model for each pictured name and the likelihood of a particular\ninstance of a name being pictured based on the surrounding words and punctuation.\n\nFace recognition cannot be surveyed reasonably in the space available. Reviews appear\nin [6, 10, 11]. Although face recognition is well studied, it does not work very well in\npractice [15]. One motivation for our work is to take the large collection of news images\nand captions as semi-supervised input and and produce a fully supervised dataset of faces\nlabeled with names. The resulting dataset exhibits many of the confounding factors that\nmake real-world face recognition difficult, in particular modes of variation that are not\nfound in face recognition datasets collected in laboratories. It is important to note that this\ntask is easier than general face recognition because each face has only a few associated\nnames.\n\n\f\n                                                                               N1                 F1       N1               F1\n\n\n                                                                               N2                 F2       N2               F2\n\n\n                                                                              Null                Null    Null              Null\n\n\n                                                    F1 = \n                                                                               N1                 F1       N1               F1\n\n\n                                                                                                  F2\n                                                                               N2                          N2               F2\n\n\n                                                                              Null                Null    Null              Null\n                                                    F2 = \n\n\n                                                                               N1                 F1       N1               F1\n\n                                                    N1 = John Furlong,         N2                 F2       N2               F2\n                                                    N2 = Winter Olympics, \n\n                                                                              Null                Null    Null              Null\n\n\n                                                                                           N1                F1\nPresident and Chief Operating Officer of the Vancouver, \nBritish Columbia 2010 Bid Corporation John Furlong                                         N2                F2\n(rear) smiles while celebrating with compatriots their\nvictory in obtaining the 2010 Winter Olympics bid on late\n                                                                                          Null               Null\nJuly 2, 2003 in Prague. Vancouver won with 56 votes \nagainst 53 votes for Pyeonchang in the second round of                                All possible name/face assignments\nballoting at an IOC gathering in Prague. REUTERS/Petr \nJosek\n\nFigure 1: Left: A typical data item consisting of a picture and its associated caption.\nCenter: Detected faces and names for this data item. Right: The set of possible corre-\nspondences for this data item. Our model allows each face to be assigned to at most one\nname, each name to be assigned to at most one face, and any face or name to be assigned\nto Null. Our named entity recognizer occasionally identifies strings that do not refer to\nactual people (e.g. \"Winter Olympics\") . These names are assigned low probability under\nour model and therefore their assignment to a face is unlikely. EM iterates between com-\nputing the expectation of the possible face-name correspondences and updating the face\nclusters and language model. Unusually, we can afford to compute all possible face-name\ncorrespondences for a data item since the number of possibilities is small. For this item,\nwe correctly choose the best matching \"F1 to Null\", \"N2 to Null\", and \"F2 to N1\".\n\nLanguage: Quite simple common phenomena in captions suggest using a language model.\nFirst, our named entity recognizer occasionally marks incorrect names like \"United Na-\ntions\". The context in which these incorrect detections occur suggest that they do not refer\nto actual people. Second, name-context pairs can be weighted according to their probabil-\nity. In a caption such as \"Michael Jackson responds to questioning Thursday, Nov. 14, 2002\nin Santa Maria Superior Court in San ta Maria, Calif., during a $21 million lawsuit brought\nagainst him by Marcel Avram for failing to appear at two millennium concerts...\", Michael\nJackson appears in a more favorable context (at the beginning of the caption, followed by\na verb) than Marcel Avram (near the middle of the caption, followed by a preposition).\n\nOur approach combines a simple appearance model using kPCA and LDA, with a lan-\nguage model, based on context. We evaluate both an EM and maximum likelihood clus-\ntering and show that incorporating language with appearance produces better results than\nusing appearance alone. We also show the results of the learned natural language classifier\napplied to a set of captions in isolation.\n\n\n2        Linking a face and language model with EM\n\nA natural way of thinking about name assignment is as a hidden variable problem where\nthe hidden variables are the correct name-face correspondences for each picture. This sug-\ngests using an expectation maximization (EM) procedure. EM iterates between computing\n\n\f\n before - CEO Summit           before - U.S. Joint       before - Angelina Jolie         before - Ric Pipino        before - U.S. Open        before - James Bond\n after - Martha Stewart        after - Null               after - Jon Voight             after - Heidi Klum       after - David Nalbandian    after - Pierce Brosnan\n\n\n\n\n\n before - U.S. House       before - Julia Vakulenko      before - Vice President       before - Marcel Avram        before - al Qaeda          before - James Ivory\nafter - Andrew Fastow      after - Jennifer Capriati             Dick Cheney           after - Michael Jackson       after - Null              after - Naomi Watts\n                                                        after - President George W.\n\nFigure 2: Names assigned using our raw clustering procedure (before) and incorporating a\nlanguage model (after). Our named entity recognizer occasionally detects incorrect names\n(e.g. \"CEO Summit\"), but based on context the language model assigns low probabilities\nto these names, making their assignment unlikely. When multiple names are detected like\n\"Julia Vakulenko\" and \"Jennifer Capriati\", the probability for each name depends on its\ncontext. The caption for this picture reads \"American Jennifer Capriati returns the ball to\nher Ukrainian opponent Julia Vakulenko in Paris during...\" The language model prefers to\nassign the name \"Jennifer Capriati\" because its context (beginning of the caption followed\nby a present tense verb) indicates it is more likely to be pictured than \"Julia Vakulenko\"\n(middle of the caption followed by a preposition). For pictures such the man labeled \"al\nQaeda\" to \"Null\" where the individual is not named in the caption, the language model\ncorrectly assigns \"Null\" to the face. As table 1 shows, incorporating a language model\nimproves our face clusters significantly.\n\n\n\n\n\nthe expected values of the set of face-name correspondences (given a face clustering and\nlanguage model) and updating the face clusters and language model given the correspon-\ndences. Unusually, it is affordable to compute the expected value of all possible face-name\ncorrespondences for a data item since the number of possibilities is small.\n\nTo use EM we need a model of how pictures are generated. Generative model:\n\n                                                                To generate a data item:\n                  name, context                                           1. Choose N, the number of names, and F, the\n                                                                                number of faces.\n                                                                          2. Generate N name, context pairs.\n                                                                          3. For each of these name, context pairs,\n                  pictured\n                                     N                                          generate a binary variable pictured con-\n                                                                                ditioned on the context alone (from\n                                                                                P (pictured|context)).\n                                                                          4. For each pictured = 1, generate a face\n     face_u                face_n\n               Fu                    Fn                                         from P (f ace|name) (F n =                                    pictured).\n\n                                           D                              5. Generate F u = F - F n other faces from\n                                                                                P (f ace).\n\n\nThe parameters necessary for the EM process are P (f ace|name) (sec 2.2), the probability\nthat a face is generated by a given name, P (pictured|context) (sec 2.3), the probability\nthat a name is pictured given its context, and P (f ace) the probability that a face is gener-\nated without a name.\n\n\f\n                                                               Elizabeth Dole\n       Angelina Jolie                                                              Angelina Jolie\n                                        Elizabeth Dole\n\n\n\n\n\n                                       Donald Fehr\n                                                               Donald Fehr\n\n\n\n\n\n      U.S. Open                                                 Abraham Lincoln\n                                         Abraham Lincoln\n                                                                                            U.S. Open\n\n                                                                   empty\n\n                                                                                                empty\n\n                                          Anastasia Myskina\n                                                                 Anastasia Myskina\n\n\n\n\n\n       Without language model                                     With language model\n\nFigure 3: Left:               Example clusters using only appearance to cluster. Right: The same\nclusters, but using appearance + language to cluster. Some clusters get larger (Elizabeth\nDole, Angelina Jolie) because of the inclusion of more correct faces. Some clusters get\nsmaller (Anastasia Myskina) because of the exclusion of incorrect faces. All clusters get\nmore accurate because the language model is breaking ambiguities and giving the cluster-\ning a push in the right direction. Some clusters that do not refer to actual people, \"U.S.\nOpen\" completely disappear using the language model. Other clusters like \"Abraham Lin-\ncoln\" (who is a person, but whose associated pictures most often portray people other than\n\"Abraham Lincoln\") become empty when using the language model, presumably because\nthese faces are assigned to the correct names.\n\n\n2.1       Name Assignment\n\nFor each image-caption pair, we calculate the costs of all possible assignments of names to\nfaces (dependent upon the associated faces and names) and use the best such assignment.\nAn example of the extracted names, faces and all possible assignments can be seen in figure\n1.\n\nThe likelihood of picture xi under assignment aj, of names to faces under our generative\nmodel is:\n\n      L(xi, aj) =P (N )P (F )P (n1, c1)...P (nn, cn)\n\n                              P (pictured|c)P (f()|n)             (1 - P (pictured|c))                 P (f)\n                                                                                                       \n\n\nWhere P(N) is the probability of generating N names, P(F) is the probability of generating F\nfaces, and P (ni, ci) are the probabilities of generating namei and context ci. In assignment\naj,  indexes into the names that are pictured, () indexes into the faces assigned to the\npictured names,  indexes into the names that are not pictured and  indexes into the faces\nwithout assigned names. The terms P (N )P (F )P (n1, c1)...P (nn, cn) are not dependent\non the assignment so we can ignore them when calculating the probability of an assignment\nand focus on the remaining terms.\n\n\f\n IN Pete Sampras IN of the U.S. celebrates his victory over Denmark's OUT Kristian Pless\n OUT at the OUT U.S. Open OUT at Flushing Meadows August 30, 2002. Sampras won the\n match 6-3 7- 5 6-4. REUTERS/Kevin Lamarque\n Germany's IN Chancellor Gerhard Schroeder IN, left, in discussion with France's IN Presi-\n dent Jacques Chirac IN on the second day of the EU summit at the European Council headquar-\n ters in Brussels, Friday Oct. 25, 2002. EU leaders are to close a deal Friday on finalizing entry\n talks with 10 candidate countries after a surprise breakthrough agreement on Thursday between\n France and Germany regarding farm spending.(AP Photo/European Commission/HO)\n 'The Right Stuff' cast members IN Pamela Reed IN, (L) poses with fellow cast member IN\n Veronica Cartwright IN at the 20th anniversary of the film in Hollywood, June 9, 2003. The\n women played wives of astronauts in the film about early United States test pilots and the space\n program. The film directed by OUT Philip Kaufman OUT, is celebrating its 20th anniversary\n and is being released on DVD. REUTERS/Fred Prouser\n Kraft Foods Inc., the largest U.S. food company, on July 1, 2003 said it would take steps, like\n capping portion sizes and providing more nutrition information, as it and other companies face\n growing concern and even lawsuits due to rising obesity rates. In May of this year, San Fran-\n cisco attorney OUT Stephen Joseph OUT, shown above, sought to ban Oreo cookies in Cali-\n fornia  a suit that was withdrawn less than two weeks later. Photo by Tim Wimborne/Reuters\n REUTERS/Tim Wimborne\n\nFigure 4: Our new procedure gives us not only better clustering results, but also a natural\nlanguage classifier which can be tested on captions in isolation. Above: a few captions\nlabeled with IN (pictured) and OUT (not pictured) using our learned language model. Our\nlanguage model has learned which contexts have high probability of referring to pictured\nindividuals and which contexts have low probabilities. We observe an 85% accuracy of\nlabeling who is portrayed in a picture using only our language model. The top 3 labelings\nare all correct. The last incorrectly labels \"Stephen Joseph\" as not pictured when in fact\nhe is the subject of the picture. Some contexts that are often incorrectly labeled are those\nwhere the name appears near the end of the caption (usually a cue that the individual\nnamed is not pictured). Some cues we could add that should improve the accuracy of our\nlanguage model are the nearness of words like \"shown\", \"pictured\", or \"photographed\".\n\nThe complete data log likelihood is:\n\n                                                                      \n                                                     (Pijlog(L(xi, aj))\n                              i pics         j ass\n\n\nWhere Pij is an indicator variable telling which correspondence occurred in this data item.\nThe Pij are missing data whose expectations are computed in the E step.\n\nThis gives a straightforward EM procedure:\n\n        E  update the Pij according to the normalized probability of picture i with as-\n          signment j.\n        M  maximize the parameters P (f ace|name) and P (pictured|context) using\n          soft counts.\n\n\n2.2    Modeling the appearance of faces  P (f ace|name)\n\nWe model appearance using a mixture model with one mixture element per name in our\nlexicon. We need a representation for faces in a feature space where comparisons are\nhelpful. Our representation is obtained by rectification of the faces followed by kernel\nprincipal components analysis (kPCA) and linear discriminant analysis (LDA) (details in\n[5]). We model the distributions P (f ace|name) using gaussians with fixed covariance.\n\nTo obtain features we first automatically rectify all faces to a canonical pose. Five support\nvector machines are trained as feature detectors (corners of the left and right eyes, corners\n\n\f\n                                    Model                          EM     MM\n                      Appearance Model, No Lang Model              56%    67%\n                        Appearance Model + Lang Model              72%    77%\n\nTable 1: Above:       We randomly selected a set of 1000 faces from our dataset and hand\nlabeled them with their correct names. Here we show what percentage of those faces are\ncorrectly labeled by each of our methods (EM and maximal correspondence clustering,\nMM). For both methods, incorporating a language model improves their respective clus-\nterings greatly. Standard statistical knowledge says that using the expected values should\nperform better than simply choosing the maximal assignment at each step (MM). However,\nwe have found that using the maximal assignment works better than taking an expectation.\nOne reason this could be true is that EM averages incorrect faces into the appearance\nmodel, making the mean unstable.\n\nof the mouth and the tip of the nose) using features consisting of the geometric blur of [4]\napplied to grayscale patches. We then use kPCA ([16]) to reduce the dimensionality of our\ndata and compute linear discriminants ([3]) on the single name, single face pictures.\n\nBecause of the large size of our dataset, we cannot compute the kernel matrix, K for kPCA,\ndirectly. Instead we use an approximation to calculate the eigenvectors of K, the Nystr om\napproximation (cf [17, 9]). The Nystr om approximation computes two exact subsets of K\nand uses these to efficiently approximate the rest of K and its eigenvectors (details in [5]).\n\n\n2.3    Language Model  P (pictured|context)\n\nOur language model assigns a probability to each name based on its context within the\ncaption. These distributions, P (pictured|context), are learned using counts of how often\neach context appears describing an assigned name, versus how often that context appears\ndescribing an unassigned name. We have one distribution for each possible context cue,\nand assume that context cues are modeled independently (because we lack enough data to\nmodel them jointly).\n\nFor context, we use a variety of cues: the part of speech tags of the word immediately\nprior to the name and immediately after the name within the caption (modeled jointly), the\nlocation of the name in the caption, and the distances to the nearest \",\", \".\", \"(\", \")\", \"(L)\",\n\"(R)\" and \"(C)\". We tried adding a variety of other language model cues, but found that\nthey did not increase the assignment accuracy.\n\nThe probability of being pictured given multiple context cues (where Ci are the different\nindependent context cues) can be formed using Bayes rule:\n\n                                                 P (pictured|C\n            P (pictured|C                                         1)...P (pictured|Cn)\n                             1, C2, ...Cn) =             P (pictured)n-1\n\nWe compute maximum likelihood estimates of each of P (pictured|Ci) and P (pictured)\nusing soft counts.\n\n\n2.4    Best correspondence and mean correspondence\n\nGiven our hidden variable problem of determining correct name-face assignments, from a\nstatistics point of view EM seems like the most favorable choice. However, many computer\nvision problems have observed better results by choosing maximum over expected values.\nWe have tried both methods and found that using the maximal assignment produced better\nresults (table 1). One reason this might be true is that for cases where there is a clear best\nassignment the max and the average are basically equivalent. For cases where there is no\nclear best, EM averages over assignments, producing a mean that has no real meaning since\nit is an average of different people's faces.\n\n\f\n                      Classifier               labels correct     IN correct    OUT correct\n                      Baseline                    67%              100%            0%\n       EM Labeling with Language Model            76%              95%            56%\n       MM Labeling with Language Model            84%              87%            76%\n\nTable 2: Above: Results of applying our learned language model to a test set of 430\ncaptions (text alone). In our test set, we hand labeled each detected name with IN/OUT\nbased on whether the referred name was pictured within the corresponding picture. We\nthen tested how well our language model could predict those labels (\"labels correct\" refers\nto the total percentage of names that were correctly labeled, \"IN correct\" the percentage\nof pictured names correctly labeled, and \"OUT correct\" the percentage of not pictured\nnames correctly labeled). The baseline figure gives the accuracy of labeling all names\nas pictured. Using EM to learn a language model gives an accuracy of 76% while using\na maximum likelihood clustering gives 84%. Again the maximum likelihood clustering\noutperforms EM. Names that are most often mislabeled are those that appear near the end\nof the caption or in other contexts that usually denote a name being not pictured.\n\nThe Maximal Assignment process is nearly the same as the EM process except instead of\ncalculating the expected value of each assignment only the maximal assignment is nonzero.\n\nThe Maximal Assignment procedure:\n          M1  set the maximal Pij to 1 and all others to 0.\n          M2  maximize the parameters P (f ace|name) and P (pictured|context) using\n           counts.\n\n3      Results\n\nWe have collected a dataset consisting of approximately half a million news pictures and\ncaptions from Yahoo News over a period of roughly two years.\n\nFaces: Using the face detector of [14], we extract 44,773 large well detected face images.\nSince these pictures were taken \"in the wild\" rather than under fixed laboratory conditions,\nthey represent a broad range of individuals, pose, expression, illumination conditions and\ntime frames. Our face recognition dataset is more varied than any other to date.\n\nNames: We use an open source named entity recognizer ([8]) to detect proper names in\neach of the associated captions. This gives us a set of names associated with each picture.\n\nScale: We obtain 44,773 large and reliable face detector responses. We reject face images\nthat cannot be rectified satisfactorily, leaving 34,623. Finally, we concentrate on images\nwithin whose captions we detect proper names, leaving 30,281, the final set we cluster on.\n\n\n3.1     Quantitative Results\n\nIncorporating a natural language model into face clustering produces much better results\nthan clustering on appearance alone. As can be seen in table 1, using, only appearance\nproduces an accuracy of 67% while appearance + language gives 77%. For face labeling,\nusing the maximum likelihood assignment (MM) rather than the average (EM) produces\nbetter results (77% vs 72%).\n\nOne neat by-product of our clustering is a natural language classifier. We can evaluate that\nclassifier on text without associated pictures. In table 2, we show results for labeling names\nwith pictured and not pictured using our language model. Using the language model we\ncorrectly label 84% of the names while the baseline (labeling everyone as pictured) only\ngives 67%. The maximum likelihood assignment also produces a better language model\nthan EM (76% vs 84%). A few things that our language model learns as indicative of being\npictured are being near the beginning of the caption, being followed by a present tense verb,\nand being near \"(L)\", \"(R)\", or \"(C)\".\n\n\f\n4     Discussion\nWe have shown previously ([5]) that a good clustering can be created using names and\nfaces. In this work, we show that by analyzing language more carefully we can produce\na much better clustering (table 1). Not only do we produce better face clusters, but we\nalso learn a natural language classifier that can be used to determine who is pictured from\ntext alone (table 2). We have coupled language and images, using language to learn about\nimages and images to learn about language.\n\nThe next step will be to try to learn a language model for free text on a webpage. One area\nwe would like to apply this to is improving google image search results. Using a simple\nimage representation and a modified context model perhaps we could link google images\nwith the words on the surrounding webpages to improve search results.\n\n\nReferences\n\n[1] K. Barnard, D.A. Forsyth, \"Clustering Art,\" Computer Vision and Pattern Recognition,\n     2001\n[2] K. Barnard, P. Duygulu, N. de Freitas, D.A. Forsyth, D. Blei, and M.I. Jordan, \"Match-\n     ing Words and Pictures,\" Journal of Machine Learning Research, Vol 3, pp 1107-1135,\n     2003.\n[3] P. Belhumeur, J. Hespanha, D. Kriegman \"Eigenfaces vs. Fisherfaces: Recognition\n     Using Class Specific Linear Projection\" Transactions on Pattern Analysis and Machine\n     Intelligence, Special issue on face recognition, pp. 711-720, July 1997.\n[4] A.C. Berg, J. Malik, \"Geometric Blur for Template Matching,\" Computer Vision and\n     Pattern Recognition,Vol I, pp. 607-614, 2001.\n[5] T.L. Berg, A.C. Berg, J. Edwards, M. Maire, R. White, E. Learned-Miller, D.A. Forsyth\n     \"Names and Faces in the News\" Computer Vision and Pattern Recognition, 2004.\n[6] V. Blanz, T. Vetter, \"Face Recognition Based on Fitting a 3D Morphable Model,\"\n     Transactions on Pattern Analysis and Machine Intelligence Vol. 25 no.9, 2003.\n[7] C. Carson, S. Belongie, H. Greenspan, J. Malik, \"Blobworld  Image segmentation us-\n     ing expectationmaximization and its application to image querying,\" IEEE Transactions\n     on Pattern Analysis and Machine Intelligence, 24(8), pp. 10261038, 2002.\n[8] H. Cunningham, D. Maynard, K. Bontcheva, V. Tablan, \"GATE: A Framework and\n     Graphical Development Environment for Robust NLP Tools and Applications,\" 40th\n     Anniversary Meeting of the Association for Computational Linguistics\", Philadelphia,\n     July 2002.\n[9] C. Fowlkes, S. Belongie, F. Chung, J. Malik, \"Spectral Grouping Using The Nystr om\n     Method,\" TPAMI, Vol. 26, No. 2, February 2004.\n[10] R. Gross, J. Shi and J. Cohn, \"Quo Vadis Face Recognition?,\" Third Workshop on\n     Empirical Evaluation Methods in Computer Vision, December, 2001.\n[11] R. Gross, I. Matthews, and S. Baker, \"Appearance-Based Face Recognition and Light-\n     Fields,\" Transactions on Pattern Analysis and Machine Intelligence, 2004.\n[12] V. Lavrenko, R. Manmatha., J. Jeon, \"A Model for Learning the Semantics of Pic-\n     tures,\" Neural Information Processing Systems, 2003\n[13] J. Li and J. Z. Wang, \"Automatic Linguistic Indexing of Pictures by a Statistical Mod-\n     eling Approach,\" Transactions on Pattern Analysis and Machine Intelligence, vol. 25,\n     no. 9, pp. 1075-1088, 2003\n[14] K. Mikolajczyk \"Face detector,\" Ph.D report, INRIA Rhone-Alpes\n[15] J.       Scheeres,    \"Airport    face    scanner    failed\",    Wired      News,    2002.\n     http://www.wired.com/news/privacy/0,1848,52563,00.html.\n[16] B. Scholkopf, A. Smola, K.-R. Muller \"Nonlinear Component Analysis as a Kernel\n     Eigenvalue Problem\" Neural Computation, Vol. 10, pp. 1299-1319, 1998.\n[17] C. Williams, M. Seeger \"Using the Nystr om Method to Speed up Kernel Machines,\"\n     Advances in Neural Information Processing Systems, Vol 13, pp. 682-688, 2001.\n\n\f\n", "award": [], "sourceid": 2708, "authors": [{"given_name": "Tamara", "family_name": "Berg", "institution": null}, {"given_name": "Alexander", "family_name": "Berg", "institution": null}, {"given_name": "Jaety", "family_name": "Edwards", "institution": null}, {"given_name": "David", "family_name": "Forsyth", "institution": null}]}