| [ |
| { |
| "chunk_id": "d34e347f-6428-4e9d-8996-2044f7413ede", |
| "text": "Machine Learning that Matters Wagstaff kiri.l.wagstaff@jpl.nasa.gov\nJet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena, CA 91109 USA", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 0, |
| "total_chunks": 28, |
| "char_count": 175, |
| "word_count": 21, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "b86e0d96-afbb-4ad0-b0ab-697349e62a80", |
| "text": "Abstract tively solved spam email detection (Zdziarski, 2005)\nand machine translation (Koehn et al., 2003), two\nMuch of current machine learning (ML) reproblems of global import. And so on.\nsearch has lost its connection to problems of\nimport to the larger world of science and so- And yet we still observe a proliferation of published\nciety. From this perspective, there exist glar- ML papers that evaluate new algorithms on a handful\ning limitations in the data sets we investi- of isolated benchmark data sets. Their \"real world\"\ngate, the metrics we employ for evaluation, experiments may operate on data that originated in\nand the degree to which results are commu- the real world, but the results are rarely communicated\nnicated back to their originating domains. back to the origin. Quantitative improvements in perWhat changes are needed to how we con- formance are rarely accompanied by an assessment of\nduct research to increase the impact that ML whether those gains matter to the world outside of\nhas? We present six Impact Challenges to ex- machine learning research.\nplicitly focus the field's energy and attention,\nThis phenomenon occurs because there is no\nand we discuss existing obstacles that must\nwidespread emphasis, in the training of graduate stube addressed. We aim to inspire ongoing disdent researchers or in the review process for submitted\ncussion and focus on ML that matters.\npapers, on connecting ML advances back to the larger\nworld. Even the rich assortment of applications-driven\nML research often fails to take the final step to trans-\n1. Introduction late results into impact. At one time or another, we all encounter a friend, Many machine learning problems are phrased in terms\nspouse, parent, child, or concerned citizen who, upon of an objective function to be optimized. It is time for\nlearning that we work in machine learning, wonders us to ask a question of larger scope: what is the field's\n\"What's it good for?\" The question may be phrased objective function? Do we seek to maximize performore subtly or elegantly, but no matter its form, it gets mance on isolated data sets? Or can we characterize\nat the motivational underpinnings of the work that we progress in a more meaningful way that measures the\ndo. Why do we invest years of our professional lives concrete impact of machine learning innovations?\nin machine learning research? What difference does it\nThis short position paper argues for a change in howmake, to ourselves and to the world at large?\nwe view the relationship between machine learning and\nMuch of machine learning (ML) research is inspired science (and the rest of society).", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 1, |
| "total_chunks": 28, |
| "char_count": 2644, |
| "word_count": 435, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "b99c9e94-ea3a-415d-b76f-4e2a1d038c75", |
| "text": "This paper does not\nby weighty problems from biology, medicine, finance, contain any algorithms, theorems, experiments, or reastronomy, etc.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 2, |
| "total_chunks": 28, |
| "char_count": 140, |
| "word_count": 19, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "39a55b38-5b10-47a3-9d48-977d270d7772", |
| "text": "The growing area of computational sults. Instead it seeks to stimulate creative thought\nsustainability (Gomes, 2009) seeks to connect ML ad- and research into a large but relatively unaddressed isvances to real-world challenges in the environment, sue that underlies much of the machine learning field.\neconomy, and society. The CALO (Cognitive Assistant The contributions of this work are 1) the clear identifithat Learns and Organizes) project aimed to integrate cation and description of a fundamental problem: the\nlearning and reasoning into a desktop assistant, poten- frequent lack of connection between machine learning\ntially impacting everyone who uses a computer (SRI research and the larger world of scientific inquiry and\nInternational, 2003–2009). Machine learning has effec- humanity, 2) suggested first steps towards addressing\nthis gap, 3) the issuance of relevant Impact Challenges\nAppearing in Proceedings of the 29 th International Confer- to the machine learning community, and 4) the idenence on Machine Learning, Edinburgh, Scotland, UK, 2012.\ntification of several key obstacles to machine learningCopyright 2012 California Institute of Technology. Machine Learning that Matters impact, as an aid for focusing future research efforts. viewers do not inquire as to which classes were well\nWhether or not the reader agrees with all statements classified and which were not, what the common erin this paper, if it inspires thought and discussion, then ror types were, or even why the particular data sets\nits purpose has been achieved. were chosen. There is no expectation that the authors report whether an observed x% improvement in\nperformance promises any real impact for the original2.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 3, |
| "total_chunks": 28, |
| "char_count": 1710, |
| "word_count": 258, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "d79bfcb6-014b-4a5d-a975-fa3c414d75cd", |
| "text": "Machine Learning for Machine\ndomain. Even when the authors have forged a colLearning's Sake laboration with qualified experts, little paper space is\nThis section highlights aspects of the way ML research devoted to interpretation, because we (as a field) do\nis conducted today that limit its impact on the larger not require it.\nworld. Our goal is not to point fingers or critique indi- The UCI archive has had a tremendous impact on the\nviduals, but instead to initiate a critical self-inspection field of machine learning. Legions of researchers have\nand constructive, creative changes. These problems do chased after the best iris or mushroom classifier. Yet\nnot trouble all ML work, but they are common enough this flurry of effort does not seem to have had any imto merit our effort in eliminating them. pact on the fields of botany or mycology. Do scientists\nThe argument here is also not about \"theory versus in these disciplines even need such a classifier? Do\napplications.\" Theoretical work can be as inspired by they publish about this subject in their journals?\nreal problems as applied work can. The criticisms here There is not even agreement in the community about\nfocus instead on the limitations of work that lies be- what role the UCI data sets serve (benchmark? \"realtween theory and meaningful applications: algorith- world\"?). They are of less utility than synthetic data,\nmic advances accompanied by empirical studies that since we did not control the process that generated\nare divorced from true impact. them, and yet they fail to serve as real world data\ndue to their disassociation from any real world con-\n2.1. Hyper-Focus on Benchmark Data Sets text (experts, users, operational systems, etc.). It is\nas if we have forgotten, or chosen to ignore, that eachIncreasingly, ML papers that describe a new algorithm\ndata set is more than just a matrix of numbers.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 4, |
| "total_chunks": 28, |
| "char_count": 1885, |
| "word_count": 312, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "1205e802-3d20-41c3-8451-3c3b9430e1c1", |
| "text": "Fur-follow a standard evaluation template. After presentther, the existence of the UCI archive has tended toing results on synthetic data sets to illustrate certain\nover-emphasize research effort on classification and re-aspects of the algorithm's behavior, the paper reports\ngression problems, at the expense of other ML prob-results on a collection of standard data sets, such as\nlems (Langley, 2011). Informal discussions with otherthose available in the UCI archive (Frank & Asuncion,\nfine features, leaving young researchers unprepared to\n148/152 (93%) include experiments of some sort tackle new problems.\n57/148 (39%) use synthetic data\n55/148 (37%) use UCI data This trend has been going on for at least 20 years.\n34/148 (23%) use ONLY UCI and/or synthetic data Jaime Carbonell, then editor of Machine Learning,\n1/148 (1%) interpret results in domain context wrote in 1992 that \"the standard Irvine data sets are\nused to determine percent accuracy of concept clas-The possible advantages of using familiar data sets insification, without regard to performance on a largerclude 1) enabling direct empirical comparisons with\nexternal task\" (Carbonell, 1992). Can we change thatother methods and 2) greater ease of interpreting\ntrend for the next 20 years? Do we want to?the results since (presumably) the data set properties\nhave been widely studied and understood. Hyper-Focus on Abstract Metricsin practice direct comparisons fail because we have\nno standard for reproducibility. Experiments vary in There are also problems with how we measure permethodology (train/test splits, evaluation metrics, pa- formance. Most often, an abstract evaluation metric\nrameter settings), implementations, or reporting.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 5, |
| "total_chunks": 28, |
| "char_count": 1712, |
| "word_count": 252, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "eb74d200-2599-4136-ae69-57ac406204ad", |
| "text": "In- (classification accuracy, root of the mean squared erterpretations are almost never made. Why is this? ror or RMSE, F-measure (van Rijsbergen, 1979), etc.)\nFirst, meaningful interpretations are hard. These metrics are abstract in that they exnone of the ML researchers who work with these data plicitly ignore or remove problem-specific details, ususets happen to also be experts in the relevant scientific ally so that numbers can be compared across domains.\ndisciplines. Second, and more insidiously, the ML field Does this seemingly obvious strategy provide us with\nneither motivates nor requires such interpretation.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 6, |
| "total_chunks": 28, |
| "char_count": 624, |
| "word_count": 93, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "8d57bd95-cc84-402f-bb93-0178f9faa032", |
| "text": "Re- useful information? Machine Learning that Matters Phrase problem Select or\nCollect Necessary\nas a machine generate\ndata\nlearning task features preparation Choose or develop Choose metrics, The \"machine\nalgorithm conduct experiments learning contribution\" Interpret Publicize results to Persuade users to Impact\nresults relevant user community adopt technique Three stages of a machine learning research program. Current publishing incentives are highly biased towards\nthe middle row only.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 7, |
| "total_chunks": 28, |
| "char_count": 492, |
| "word_count": 67, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "1252e5ad-d65e-46ff-846d-26b148ac3715", |
| "text": "It is recognized that the performance obtained by Methods from statistics such as the t-test (Student,\ntraining a model M on data set X may not reflect M's 1908) are commonly used to support a conclusion\nperformance on other data sets drawn from the same about whether a given performance improvement is\nproblem, i.e., training loss is an underestimate of test \"significant\" or not. Statistical significance is a funcloss (Hastie et al., 2001). Strategies such as splitting tion of a set of numbers; it does not compute real-world\nX into training and test sets or cross-validation aim significance. Of course we all know this, but it rarely\nto estimate the expected performance of M′, trained inspires the addition of a separate measure of (true)\non all of X, when applied to future data X′. significance. How often, instead, a t-test result serves\nas the final punctuation to an experimental utterance!\nHowever, these metrics tell us nothing about the impact of different performance. For example, 80% ac-\n2.3.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 8, |
| "total_chunks": 28, |
| "char_count": 1011, |
| "word_count": 167, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "98375f45-f439-41f1-945b-009b45eeff23", |
| "text": "Lack of Follow-Throughcuracy on iris classification might be sufficient for the\nbotany world, but to classify as poisonous or edible It is easy to sit in your office and run a Weka (Hall\na mushroom you intend to ingest, perhaps 99% (or et al., 2009) algorithm on a data set you downloaded\nhigher) accuracy is required. The assumption of cross- from the web. It is very hard to identify a problem\ndomain comparability is a mirage created by the appli- for which machine learning may offer a solution, decation of metrics that have the same range, but not the termine what data should be collected, select or exsame meaning. Suites of experiments are often sum- tract relevant features, choose an appropriate learning\nmarized by the average accuracy across all data sets. method, select an evaluation method, interpret the reThis tells us nothing at all useful about generalization sults, involve domain experts, publicize the results to\nor impact, since the meaning of an x% improvement the relevant scientific community, persuade users to\nmay be very different for different data sets. A related adopt the technique, and (only then) to truly have\nproblem is the persistence of \"bake-offs\" or \"mind- made a difference (see Figure 1).", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 9, |
| "total_chunks": 28, |
| "char_count": 1232, |
| "word_count": 204, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "7a58fcbd-0996-4524-86f8-27f28174dfd1", |
| "text": "An ML researcher\nless comparisons among the performance of algorithms might well feel fatigued or daunted just contemplating\nthat reveal little about the sources of power or the ef- this list of activities. However, each one is a necessary\nfects of domain characteristics\" (Langley, 2011). component of any research program that seeks to have\na real impact on the world outside of machine learning.Receiver Operating Characteristic (ROC) curves are\nused to describe a system's behavior for a range of Our field imposes an additional obstacle to impact.\nthreshold settings, but they are rarely accompanied Generally speaking, only the activities in the middle\nby a discussion of which performance regimes are rele- row of Figure 1 are considered \"publishable\" in the ML\nvant to the domain. The common practice of reporting community. The Innovative Applications of Artificial\nthe area under the curve (AUC) (Hanley & McNeil, Intelligence conference and International Conference\n1982) has several drawbacks, including summarizing on Machine Learning and Applications are exceptions.\nperformance over all possible regimes even if they are The International Conference on Machine Learning\nunlikely ever to be used (e.g., extremely high false pos- (ICML) experimented with an (unreviewed) \"invited\nitive rates), and weighting false positives and false neg- applications\" track in 2010. Yet to be accepted as a\natives equally, which may be inappropriate for a given mainstream paper at ICML or Machine Learning or\nproblem domain (Lobo et al., 2008). As such, it is in- the Journal of Machine Learning Research, authors\nsufficiently grounded to meaningfully measure impact. must demonstrate a \"machine learning contribution\"\nthat is often narrowly interpreted by reviewers as \"the Machine Learning that Matters development of a new algorithm or the explication of a system perfectly solves a sub-problem of little interest\nnovel theoretical analysis.\" While these are excellent, to the relevant scientific community, or where the ML\nlaudable advances, unless there is an equal expectation system's performance appears good numerically but is\nfor the bottom row of Figure 1, there is little incentive insufficiently reliable to ever be adopted.\nto connect these advances with the outer world. We could also solicit short \"Comment\" papers, to acReconnecting active research to relevant real-world company the publication of a new ML advance, that\nproblems is part of the process of maturing as a re- are authored by researchers with relevant domain exsearch field. To the rest of the world, these are the pertise but who were uninvolved with the ML research.\nonly visible advances of ML, its only contributions to They could provide an independent assessment of the\nthe larger realm of science and human endeavors. performance, utility, and impact of the work. As\nan additional benefit, this informs new communities\nabout how, and how well, ML methods work. Making Machine Learning Matter\ning awareness, interest, and buy-in from ecologists,\nRather than following the letter of machine learning, astronomers, legal experts, doctors, etc., can lead to\ncan we reignite its spirit? This is not simply a matter greater opportunities for machine learning impact.\nof reporting on isolated applications. What is needed\nis a fundamental change in how we formulate, attack, 3.3.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 10, |
| "total_chunks": 28, |
| "char_count": 3360, |
| "word_count": 518, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "5de14c16-8d78-4654-9eb9-a1b0a7f71ba4", |
| "text": "Eyes on the Prize\nand evaluate machine learning research projects. Finally, we should consider potential impact when selecting which research problems to tackle, not merely\n3.1. Meaningful Evaluation Methods\nhow interesting or challenging they are from the ML\nThe first step is to define or select evaluation meth- perspective. How many people, species, countries, or\nods that enable direct measurement, wherever possi- square meters would be impacted by a solution to the\nble, of the impact of ML innovations. In addition to problem? What level of performance would constitute\ntraditional measures of performance, we can measure a meaningful improvement over the status quo?\ndollars saved, lives preserved, time conserved, effort\nWarrick et al. (2010) provides an example of ML work\nreduced, quality of living increased, and so on.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 11, |
| "total_chunks": 28, |
| "char_count": 832, |
| "word_count": 128, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "55f5e229-cc99-422d-aeae-ba20eaab5e6d", |
| "text": "Focusthat tackles all three aspects. Working with doctors\ning our metrics on impact will help motivate upstream\nand clinicians, they developed a system to detect fetal\nrestructuring of research efforts. They will guide how\nhypoxia (oxygen deprivation) and enable emergency\nwe select data sets, structure experiments, and define\nintervention that literally saves babies from brain inobjective functions.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 12, |
| "total_chunks": 28, |
| "char_count": 402, |
| "word_count": 57, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "dd4d0ed8-941e-4e0e-9f3d-31884e81236e", |
| "text": "At a minimum, publications can\njuries or death. After publishing their results, which\nreport how a given improvement in accuracy translates\ndemonstrated the ability to have detected 50% of feto impact for the originating problem domain.\ntal hypoxia cases early enough for intervention, with\nThe reader may wonder how this can be accomplished, an acceptable false positive rate of 7.5%, they are curif our goal is to develop general methods that ap- rently working on clinical trials as the next step toply across domains. Yet (as noted earlier) the com- wards wide deployment. Many such examples exist.\nmon approach of using the same metric for all do- This paper seeks to inspire more.\nmains relies on an unstated, and usually unfounded,\nassumption that it is possible to equate an x% im- 4. Machine Learning Impact Challenges\nprovement in one domain with that in another. Instead, if the same method can yield profit improve- One way to direct research efforts is to articulate amments of $10,000 per year for an auto-tire business as bitious and meaningful challenges. In 1992, Carbonell\nwell as the avoidance of 300 unnecessary surgical in- articulated a list of challenges for the field, not to interventions per year, then it will have demonstrated a crease its impact but instead to \"put the fun back into\npowerful, wide-ranging utility. machine learning\" (Carbonell, 1992).", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 13, |
| "total_chunks": 28, |
| "char_count": 1381, |
| "word_count": 225, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "aef55f72-5b43-4d52-90b5-532d7a015172", |
| "text": "Discovery of a new physical law leading to a pub-\n3.2. Involvement of the World Outside ML lished, referred scientific article. Many ML investigations involve domain experts as col- 2. Improvement of 500 USCF/FIDE chess rating\nlaborators who help define the ML problem and label points over a class B level start.\ndata for classification or regression tasks. Improvement in planning performance of 100 foldalso provide the missing link between an ML perforin two different domains.mance plot and its significance to the problem domain. This can help reduce the number of cases where an ML 4. Investment earnings of $1M in one year.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 14, |
| "total_chunks": 28, |
| "char_count": 631, |
| "word_count": 103, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "31f5a625-a38b-4e10-860b-3bb27d5ff028", |
| "text": "Machine Learning that Matters Outperforming a hand-built NLP system on a 5. Obstacles to ML Impact\ntask such as translation. Let us imagine a machine learning researcher who is\n6. Outperforming all hand-built medical diagnosis\nmotivated to tackle problems of widespread interest\nsystems with an ML solution that is deployed and\nand impact. What obstacles to success can we foresee?\nregularly used at at least two institutions. Can we set about eliminating them in advance? Because impact was not the guiding principle, these Jargon. This issue is endemic to all specialized rechallenges range widely along that axis.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 15, |
| "total_chunks": 28, |
| "char_count": 616, |
| "word_count": 97, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "b0490aca-992b-4bec-8bf8-f6e1558574dd", |
| "text": "An improved search fields. Our ML vocabulary is so familiar that it\nchess player might arguably have the lowest real-world is difficult even to detect when we're using a specialimpact, while a medical diagnosis system in active use ized term. Consider a handful of examples: \"feature\ncould impact many human lives. extraction,\" \"bias-variance tradeoff,\" \"ensemble methWe therefore propose the following six Impact Chal- ods,\" \"cross-validation,\" \"low-dimensional manifold,\"\nlenges as examples of machine learning that matters: \"regularization,\" \"mutual information,\" and \"kernel\nmethods.\" These are all basic concepts within ML\n1. A law passed or legal decision made that relies on that create conceptual barriers when used glibly to\nthe result of an ML analysis. communicate with others. Terminology can serve as\n2. $100M saved through improved decision making a barrier not just for domain experts and the general\nprovided by an ML system. public but even between closely related fields such as\n3. A conflict between nations averted through high- ML and statistics (van Iterson et al., 2012). We should\nquality translation provided by an ML system. explore and develop ways to express the same ideas in\n4. A 50% reduction in cybersecurity break-ins more general terms, or even better, in terms already\nthrough ML defenses. familiar to the audience. For example, \"feature ex-\n5.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 16, |
| "total_chunks": 28, |
| "char_count": 1379, |
| "word_count": 214, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "1c801463-01b8-4c97-b5be-54800c661740", |
| "text": "A human life saved through a diagnosis or inter- traction\" can be termed \"representation;\" the notion\nvention recommended by an ML system. of \"variance\" can be \"instability;\" \"cross-validation\"\n6. Improvement of 10% in one country's Human De- is also known as \"rotation estimation\" outside of ML;\nvelopment Index (HDI) (Anand & Sen, 1994) at- \"regularization\" can be explained as \"choosing simpler\ntributable to an ML system. models;\" and so on. These terms are not as precise,\nbut more likely to be understood, from which a conThese challenges seek to capture the entire process\nversation about further subtleties can ensue.\nof a successful machine learning endeavor, including\nperformance, infusion, and impact. They differ from Risk. Even when an ML system is no more, or less,\nexisting challenges such as the DARPA Grand Chal- prone to error than a human performing the same\nlenge (Buehler et al., 2007), the Netflix Prize (Bennett task, relying on the machine can feel riskier because\n& Lanning, 2007), and the Yahoo! Learning to Rank it raises new concerns. When errors are made, where\nChallenge (Chapelle & Chang, 2011) in that they do do we assign culpability? What level of ongoing comnot focus on any single problem domain, nor a par- mitment do the ML system designers have for adjustticular technical capability.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 17, |
| "total_chunks": 28, |
| "char_count": 1324, |
| "word_count": 215, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "001e2dbb-a124-4b02-b07c-72b717fd7f6a", |
| "text": "The goal is to inspire the ments, upgrades, and maintenance? These concerns\nfield of machine learning to take the steps needed to are especially acute for fields such as medicine, spacemature into a valuable contributor to the larger world. craft, finance, and real-time systems, or exactly those\nsettings in which a large impact is possible. An inNo such list can claim to be comprehensive, including creased sphere of impact naturally also increases the\nthis one. It is hoped that readers of this paper will associated risk, and we must address those concerns\nbe inspired to formulate additional Impact Challenges (through technology, education, and support) if we\nthat will benefit the entire field. hope to infuse ML into real systems. Much effort is often put into chasing after goals in Complexity. Despite the proliferation of ML toolwhich an ML system outperforms a human at the same boxes and libraries, the field has not yet matured to\ntask. The Impact Challenges in this paper also differ a point where researchers from other areas can simply\nfrom that sort of goal in that human-level performance apply ML to the problem of their choice (as they might\nis not the gold standard. What matters is achieving do with methods from physics, math, mechanical enperformance sufficient to make an impact on the world. gineering, etc.). Attempts to do so often fail due to\nAs an analogy, consider a sick child in a rural setting. lack of knowledge about how to phrase the problem,\nA neighbor who runs two miles to fetch the doctor what features to use, how to search over parameters,\nneed not achieve Olympic-level running speed (perfor- etc. (i.e., the top row of Figure 1). For this reason, it\nmance), so long as the doctor arrives in time to address has been said that ML solutions come \"packaged in a\nthe sick child's needs (impact).", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 18, |
| "total_chunks": 28, |
| "char_count": 1838, |
| "word_count": 312, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "070856e4-81c0-4045-8724-da55703b585a", |
| "text": "Ph.D.\"; that is, it requires the sophistication of a gradMachine Learning that Matters uate student or beyond to successfully deploy ML to Chapelle, Olivier and Chang, Yi. Yahoo! Learning to\nsolve real problems—and that same Ph.D. is needed to Rank Challenge overview. JMLR: Workshop and\nmaintain and update the system after its deployment. Conference Proceedings, 14, 2011. It is evident that this strategy does not scale to the Frank, A. and Asuncion, A.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 19, |
| "total_chunks": 28, |
| "char_count": 456, |
| "word_count": 73, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "4e4443c2-361d-45c9-9e14-cbae4e6f66e5", |
| "text": "UCI machine learning\ngoal of widespread ML impact. Simplifying, matur- repository, 2010. URL http://archive.ics.uci.\ning, and robustifying ML algorithms and tools, while edu/ml.\nitself an abstract activity, can help erode this obstacle\nGomes, Carla P. Computational sustainability: Comand permit wider, independent uses of ML.\nputational methods for a sustainable environment,\neconomy, and society. The Bridge, 39(4):5–13, Win-\n6.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 20, |
| "total_chunks": 28, |
| "char_count": 430, |
| "word_count": 58, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "f6cd163d-be08-4936-a5de-95d4167fe781", |
| "text": "Conclusions ter 2009. National Academy of Engineering. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reute-Machine learning offers a cornucopia of useful ways to\nmann, P., and Witten, I. The WEKA data min-approach problems that otherwise defy manual soluing software: An update. SIGKDD Explorations, 11tion. However, much current ML research suffers from\n(1):10–18, 2009.a growing detachment from those real problems. Many\ninvestigators withdraw into their private studies with Hanley, J.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 21, |
| "total_chunks": 28, |
| "char_count": 491, |
| "word_count": 70, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "cad561bf-a165-4b87-a06a-b8831936cb27", |
| "text": "The meaning and use\na copy of the data set and work in isolation to perfect of the area under a receiver operating characteristic\nalgorithmic performance. Publishing results to the ML (ROC) curve. Radiology, 143:29–36, 1982.\ncommunity is the end of the process. Successes usually Hastie, T., Tibshirani, R., and Friedman, J. The Elare not communicated back to the original problem ements of Statistical Learning: Data Mining, Infersetting, or not in a form that can be used. ence, and Prediction. Yet these opportunities for real impact are widespread. Koehn, Philipp, Och, Franz Josef, and Marcu, Daniel.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 22, |
| "total_chunks": 28, |
| "char_count": 605, |
| "word_count": 96, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "e6ac81e2-36be-440b-af60-c6ee304b964c", |
| "text": "The worlds of law, finance, politics, medicine, edu- Statistical phrase-based translation. In Proc. of the\ncation, and more stand to benefit from systems that Conf. of the North American Chapter of the Assocican analyze, adapt, and take (or at least recommend) ation for Computational Linguistics on Human Lanaction. This paper identifies six examples of Impact guage Technology, pp. 48–54, 2003. Challenges and several real obstacles in the hope of Langley, Pat. The changing science of machine learninspiring a lively discussion of how ML can best make ing. Machine Learning, 82:275–279, 2011.\na difference. Aiming for real impact does not just\nLobo, Jorge M., Jimnez-Valverde, Alberto, and Real,\nincrease our job satisfaction (though it may well do\nRaimundo. AUC: a misleading measure of the perthat); it is the only way to get the rest of the world to\nformance of predictive distribution models.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 23, |
| "total_chunks": 28, |
| "char_count": 899, |
| "word_count": 142, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "bdb5d0b5-2f4d-4f65-98fd-d9dcac54a58f", |
| "text": "Global\nnotice, recognize, value, and adopt ML solutions. Ecology and Biogeography, 17(2):145–151, 2008. CALO: Cognitive assistant that\nAcknowledgments learns and organizes. http://caloproject.sri. We thank Tom Dietterich, Terran Lane, Baback com, 2003–2009. Moghaddam, David Thompson, and three insightful Student.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 24, |
| "total_chunks": 28, |
| "char_count": 314, |
| "word_count": 38, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "a4782f01-8d68-47f1-9e0d-1b695e619ee9", |
| "text": "The probable error of a mean. Biometrika, 6\nanonymous reviewers for suggestions on this paper. (1):1–25, 1908. This work was performed while on sabbatical from the van Iterson, M., van Haagen, H.H.B.M., and Goeman,\nJet Propulsion Laboratory.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 25, |
| "total_chunks": 28, |
| "char_count": 241, |
| "word_count": 37, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "e692470b-1095-4553-82bb-e56da2cd2146", |
| "text": "Resolving confusion of tongues in statistics and\nmachine learning: A primer for biologists and bioinReferences formaticians. Proteomics, 12:543–549, 2012.\nvan Rijsbergen, C. Information Retrieval. Butter-Anand, Sudhir and Sen, Amartya K. Human developworth, 2nd edition, 1979. ment index: Methodology and measurement. Human Development Report Office, 1994. A machine learning approach to the de-Bennett, James and Lanning, Stan. The Netflix Prize.\ntection of fetal hypoxia during labor and delivery.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 26, |
| "total_chunks": 28, |
| "char_count": 499, |
| "word_count": 67, |
| "chunking_strategy": "semantic" |
| }, |
| { |
| "chunk_id": "7e279f07-d605-4b5a-a408-444dfee25258", |
| "text": "In In Proc. of KDD Cup and Workshop, pp. 3–6, 2007. Proc. of the Twenty-Second Innovative Applications\nBuehler, Martin, Iagnemma, Karl, and Singh, Sanjiv of Artificial Intelligence Conf., pp. 1865–1870, 2010.\n(eds.). The 2005 DARPA Grand Challenge: The\nZdziarski, Jonathan A. Ending Spam: Bayesian Con- Great Robot Race. Springer, 2007.\ntent Filtering and the Art of Statistical Language\nCarbonell, Jaime. Machine learning: A maturing field. No Starch Press, San Francisco, 2005. Machine Learning, 9:5–7, 1992.", |
| "paper_id": "1206.4656", |
| "title": "Machine Learning that Matters", |
| "authors": [ |
| "Kiri Wagstaff" |
| ], |
| "published_date": "2012-06-18", |
| "primary_category": "cs.LG", |
| "arxiv_url": "http://arxiv.org/abs/1206.4656v1", |
| "chunk_index": 27, |
| "total_chunks": 28, |
| "char_count": 510, |
| "word_count": 75, |
| "chunking_strategy": "semantic" |
| } |
| ] |
