The IBM Innovation Feed

*Originally published January 29, 2019; updated February 15, 2019, to reflect important contributions from Joy Buolamwini and Timnit Gebru in Gender Shades (2018) cited in the Diversity in Faces arXiv paper. Have you ever been treated unfairly? How did it make you feel? Probably not too good. Most people generally agree that a fairer world is a better world, and our AI researchers couldn’t agree more. That’s why we are harnessing the power of science to create AI systems that are more fair and accurate. Many of our recent advances in AI have produced remarkable capabilities for computers to accomplish increasingly sophisticated and important tasks, like translating speech across languages to bridge communications across cultures, improving complex interactions between people and machines, and automatically recognizing contents of video to assist in safety applications. Much of the power of AI today comes from the use of data-driven deep learning to train increasingly accurate models by using growing amounts of data. However, the strength of these techniques can also be a weakness. The AI systems learn what they’re taught, and if they are not taught with robust and diverse datasets, accuracy and fairness could be at risk. For that reason, IBM, along with AI developers and the research community, need to be thoughtful about what data we use for training. IBM remains committed to developing AI systems to make the world more fair. The challenge in training AI is manifested in a very apparent and profound way with facial recognition technology. There can be difficulties in making facial recognition systems that meet fairness expectations. As shown by Joy Buolamwini and Timnit Gebru in Gender Shades in 2018, facial recognition systems in commercial use performed better for lighter individuals and males and worse for darker females [1]. The heart of the problem is not with the AI technology itself, per se, but with how the AI-powered facial recognition systems are trained. For the facial recognition systems to perform as desired – and the outcomes to become increasingly accurate –  training data must be diverse and offer a breadth of coverage, as shown in our prior work [2]. For example, the training data sets must be large enough and different enough that the technology learns all the ways in which faces differ to accurately recognize those differences in a variety of situations. The images must reflect the distribution of features in faces we see in the world. How do we measure and ensure diversity for human faces? On one hand, we are familiar with how faces differ by age, gender, and skin tone, and how different faces can vary across some of these dimensions. Much of the focus on facial recognition technology has been on how well it performs within these attributes. But, as prior studies have shown, these attributes are just a piece of the puzzle and not entirely adequate for characterizing the full diversity of human faces. Dimensions like face symmetry, facial contrast, the pose the face is in, the length or width of the face’s attributes (eyes, nose, forehead, etc.) are also important. Today, IBM Research is releasing a new large and diverse dataset called Diversity in Faces (DiF) to advance the study of fairness and accuracy in facial recognition technology. The first of its kind available to the global research community, DiF provides a dataset of annotations of 1 million human facial images. Using publicly available images from the YFCC-100M Creative Commons data set, we annotated the faces using 10 well-established and independent coding schemes from the scientific literature [3-12]. The coding schemes principally include objective measures of human faces, such as craniofacial features, as well as more subjective annotations, such as human-labeled predictions of age and gender. We believe by extracting and releasing these facial coding scheme annotations on a large dataset of 1 million images of faces, we will accelerate the study of diversity and coverage of data for AI facial recognition systems to ensure more fair and accurate AI systems. Today’s release is simply the first step. We believe the DiF dataset and its 10 coding schemes offer a jumping-off point for researchers around the globe studying the facial recognition technology. The 10 facial coding methods include craniofacial (e.g., head length, nose length, forehead height), facial ratios (symmetry), visual attributes (age, gender), and pose and resolution, among others. These schemes are some of the strongest identified by the scientific literature, building a solid foundation to our collective knowledge. Our initial analysis has shown that the DiF dataset provides a more balanced distribution and broader coverage of facial images compared to previous datasets. Furthermore, the insights obtained from the statistical analysis of the 10 initial coding schemes on the DiF dataset has furthered our own understanding of what is important for characterizing human faces and enabled us to continue important research into ways to improve facial recognition technology. The dataset is available today to the global research community upon request. IBM is proud to make this available and our goal is to help further our collective research and contribute to creating AI systems that are more fair. While IBM Research is committed to continuing study and investigation of fairer facial recognition systems, we don’t believe we can do it alone. With today’s release, we urge others to contribute to the growing body of research and advance this important scientific agenda. [1]  J. Buolamwini & T. Gebru, “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,” Proc. of Machine Learning Research. 2018. [2] R. Puri, “Mitigating Bias in AI Models”, February 6, 2018. [3]  L. G. Farkas, Anthropometry of the Head and Face, Raven Press, 1994. [4]  A. Chardon I. Cretois and C. Hourseau, “Skin colour typology and suntanning pathways,” International Journal of Cosmetic Science, Aug. 1991, 13(4), pp. 191-208. [5]   Y. Liu, K. L. Schmidt, J. F. Cohn, S. Mitra, “Facial asymmetry quantification for expression invariant human identification,” Computer Vision and Image Understanding, Volume 91, Issues 1–2, July–August 2003, pp. 138-159. [6]  L. G. Farkas, et. al, “International anthropometric study of facial morphology in various ethnic groups/races,” J Craniofac Surg. 2005 Jul;16(4), pp. 615-46. [7]  N. Ramanathan, R. Chellappa, “Modeling Age Progression in Young Faces,” Intl. Conf. on Computer Vision and Pattern Recognition (CVPR), 2006, pp. 387-394. [8]  A. C. Little, B. C. Jones, L. M. DeBruine, “Facial attractiveness: evolutionary based research,” Philos Trans R Soc Lond B Biol Sci. 2011 Jun 12;366(1571), pp. 1638-59. [9]   X. Zhu, D. Ramanan, “Face Detection, Pose Estimation, and Landmark Localization in the Wild,” Intl. Conf. on Computer Vision and Pattern Recognition (CVPR), 2012, pp. 2879-2886. [10]   A. Porcheron, E. Mauger, R. Russell, “Aspects of Facial Contrast Decrease with Age and Are Cues for Age Perception,” PLoS One 8(3), Mar. 6, 2013 [11]   Z. Liu, P. Luo, X. Wang, X. Tang, “Deep Learning Face Attributes in the Wild”, Intl. Conf. on Computer Vision (ICCV), 2015, pp. 3730-3738. [12]  R. Rothe, R. Timofte, L. Van Gool, “Deep Expectation of Real and Apparent Age from a Single Image Without Facial Landmarks”, Intl. Journal of Computer Vision, Volume 126 Issue 2-4, April 2018, pp. 144-157. Researchers from MIT and IBM propose an efficient and effective method for certifying attack resistance of convolutional neural networks to given input data.

Within the next five years, the Earth’s population will cross the eight billion mark for the first time. Our complex food chain—already stressed by climate change and a finite water supply—will only be tested further. To meet the demands of this crowded future, we will need new technologies and devices, scientific breakthroughs, and entirely new ways of thinking about food safety and security. IBM researchers around the world are already working on solutions at every stage of the food chain. They are helping farmers maximize crop yields and developing ways to curb the epidemic of waste that destroys 45 percent of our food supply. Our scientists are working to create a safety net to catch pathogens and contaminants before they make people sick. And they’re inventing ways to keep plastic out of our landfills and oceans. Later this week, we’ll introduce the scientists behind this year’s 5 in 5 at a Science Slam held at the site of IBM’s biggest client event of the year: Think 2019 in San Francisco. Watch it live on Wednesday, February 13 from 10 – 11 am Pacific time, or catch the replay here. Science Slams give our researchers the opportunity to convey the importance of their work to a general audience in a very short span of time — approximately five minutes. We’ve found this to be an extremely useful exercise that makes our innovation more accessible by distilling it down to its core essentials. Our researchers inspire us to imagine what else could be possible five years from now. When the eight billionth person is born on Earth, she will enter a world more connected, more interdependent and more responsive to change than the one her parents ever imagined.  This is the future that awaits us all. How do you give a farmer who has never set foot in a bank access to credit? By digitizing and capturing all aspects of agriculture, from the quality of the soil to the skills of the tractor driver to the price of melon sold at the market. It’s known as a Digital Twin, and within the next five years, using AI we can use this data to accurately forecast crop yields, which in turn will give banks and financial institutions the data points they need to provide credit to help farmers expand — maybe money does grow on trees after all. Within five years, we’ll eliminate many of the costly unknowns in the food supply chain. From farmers to grocery suppliers, each participant in the supply chain will know exactly how much to plant, order, and ship. Food loss will diminish greatly and the produce that ends up in consumers’ carts will be fresher—when blockchain technology, IoT devices, and AI algorithms join forces. Within five years, food safety inspectors around the world will gain a new superpower: the ability to use millions of microbes to protect what we eat. These microbes—some healthy for human consumption, others not—are regularly introduced into foods at farms, factories, and grocery stores. Thanks to a new technique that enables us to analyze their genetic make-up cost effectively, microbes will tell us a lot about the safety of what we consume. Within five years, the world’s farmers, food processors, and grocers—along with its billions of home cooks—will be able to detect dangerous contaminants effortlessly in their food. All they’ll need is a cell phone or a countertop with AI sensors. IBM researchers are creating powerful, portable AI sensors that can detect foodborne pathogens anywhere and everywhere they might turn up. These mobile bacteria sensors could dramatically increase the speed of a pathogen test from days to seconds, allowing individuals up and down the food chain to detect the existence of harmful E. coli or Salmonella before it becomes an outbreak. In five years, the disposal of trash and the creation of new plastics will be completely transformed. Everything from milk cartons to cookie containers to grocery bags and cheese cloths will be recyclable, and polyester manufacturing companies will be able to take in refuse and turn it into something useful. This transition will be powered by innovations like VolCat, a catalytic chemical process that digests certain plastics (called polyesters) into a substance that can be fed directly back into plastic manufacturing machines in order to make new products.

The more AI agents are deployed in scenarios with possibly unexpected situations, the more they need to be flexible, adaptive, and creative in achieving their goals. Thus, a certain level of freedom to choose the best path to a specific goal is necessary in making AI robust and flexible enough to be deployed successfully in real-life scenarios. This is especially true when AI systems tackle difficult problems whose solution cannot be accurately defined by a traditional rule-based approach but require the data-driven and/or learning approaches increasingly being used in AI. Indeed, data-driven AI systems, such as those using machine learning, are very successful in terms of accuracy and flexibility, and they can be very “creative” in solving a problem, finding solutions that could positively surprise humans and teach them innovative ways to resolve a challenge. However, creativity and freedom without boundaries can sometimes lead to undesired actions: the AI system could achieve its goal in ways that are not considered acceptable according to values and norms of the impacted community. Thus, there is a growing need to understand how to constrain the actions of an AI system by providing boundaries within which the system must operate. This is usually referred to as the “value alignment” problem, since such boundaries should model values and principles required for the specific AI application scenario. The paper that describes our overall approach and the two possible ways to solve the value alignment problem is going to be presented at the upcoming AAAI 2019 conference and will receive the AAAI 2019 Blue Sky Idea award [1]. It can be found here. This work is part of a long-term effort to understand how to embed ethical principles into AI systems in collaboration with MIT. While the research done in [2] and [3] models ethical priorities as deontologic constraints, the IBM-MIT team is currently gathering human preferences data to model how humans follow, and switch between, different ethical theories (such as utilitarian, deontologic, and contractualist), in order to then engineer both ethical theories and switching mechanisms, suitably adapted, into AI systems. In this way, such systems will be able to be better aligned to the way people reason and act upon ethics while making decisions, and thus will be better equipped to naturally and compactly interact with humans in an augmented intelligence approach to AI. Researchers from MIT and IBM propose an efficient and effective method for certifying attack resistance of convolutional neural networks to given input data.

Throughout life, many of us develop unhealthy habits that may feel nearly impossible to change. To quit smoking, reduce alcohol consumption, eat a healthier diet, or become more physically active requires effort and the right state of mind. A team of behavioural scientists at University College London (UCL)  and researchers at IBM Research-Ireland are looking at ways to help people reach these goals and achieve better health behaviour by using AI. It is challenging to find the right ‘active ingredients’, known as Behaviour Change Techniques (BCTs), to motivate individuals towards healthier habits, help them self-manage their behaviour, and select and create environments conducive to supporting these processes. Examples of such techniques are goal setting, action planning, and self-monitoring. Policymakers are trying to find ways to develop effective programs delivered at individual, community, and population levels. What if AI could be used to help find the best BCTs for a given scenario—ultimately helping us to give up unhealthy habits and develop and sustain healthier behaviours? To help solve these types of real world problem and assist individuals, practitioners and policymakers to find the best information to help them develop effective interventions and policies, behavioural scientists and system architects at UCL have teamed up with IBM researchers in Dublin to use AI—in particular, natural language processing (NLP) and machine learning—to extract information from behaviour change literature and create an AI system to support the decision-making process in choosing interventions. Building on the internationally recognized expertise from UCL’s Centre for Behaviour Change and on IBM Research’s leading expertise in AI and NLP, the Human Behaviour Change Project (HBCP) is creating an open-access online ‘Knowledge System’ aimed at helping all those wanting to better understand evidence on behavioural science and make better decisions (e.g., policymakers, practitioners, and researchers). The project is funded by the Wellcome Trust and is a collaboration between computer scientists, behavioural scientists, and system architects. The Knowledge System that is being created will search publication databases to find behaviour change intervention evaluation reports, extract and synthesise the findings, provide answers to questions, and draw inferences about behaviour change which will enable answers to questions for which there is little direct evidence. UCL team (clockwise from top left): Mike Kelly, Emma Norris, Alison Wright, Robert West, James Thomas, Ailbhe Finnerty, Gillian Stokes, Susan Michie, Candice Moore, Marie Johnston, Marta Marques, and John Shawe-Taylor. UCL behavioural scientists are creating a Behaviour Change Intervention Ontology (BCIO) to organise the evidence using standardised terms for describing key concepts and their relationships and annotating published intervention reports using this ontology. IBM researchers are building algorithms to search, predict, recommend, and explain behaviour change interventions that are likely to be effective for a given scenario. These algorithms are being trained by the annotated research papers. As more manually annotated research papers are used to train the Knowledge System, the algorithms will become more accurate and reliable, and human input will be less needed. The third part of the collaboration is led by the system architects, who are building an online interface to make this behaviour change information easily accessible to humans and other computer programs. These three streams of work are highly iterative with ongoing interaction between the behavioural scientists, computer scientists and system architects. Both unsupervised and supervised information extraction algorithms are being used to extract information from research reports. The algorithms are defined for each information type based on a common framework. In the unsupervised setting, for each entity, we define a query used to identify passages of text likely to contain the target value (e.g., for target value = age of participants, passages must contain the word participant and age/year/old and must include an integer). Candidate answers are then extracted from each passage based on defined criteria (e.g., for age of participant the candidate answers must be integers) and re-ranked according to their alignment with the query. Rankings are proximity-based (i.e., candidate answers which are found in the text closer to the words used in the query are ranked more highly). The highest ranking answer is selected. In the supervised setting, the query is learnt automatically by a classifier trained on the articles manually annotated by the behavioural science team. These annotations are also used for evaluation. Flow of the Knowledge System. From research articles (on the left), the system extracts information guided by the Behaviour Change Intervention Ontology (BCIO). (click to enlarge) Today IBM Research-Ireland is releasing the first open-source version of their information extraction system on Github (https://github.com/HumanBehaviourChangeProject) together with a Swagger API interface. While the code itself can be used programmatically by developers interested in building systems for information extraction, the swagger interface can also be used to easily extract information from published papers reporting behaviour change evaluations. This version 0.1 of the information extraction system provides an API for extracting a limited set of data, so ‘proof of principle’. For example, user will be able to upload a pdf of a paper and ask the system to identify which of a set of 10 BCTs were used, some characteristics of the study population (e.g., minimum age, maximum age, mean age), and the outcomes and effects of the interventions. Grounded in machine learning, the work developed by the IBM Research-Ireland team can also be generalized to other kinds of applications like extracting and reasoning on information for computer science or chemistry papers, as well as to other forms of scientific productions. While we are still continuing to advance the capabilities of our algorithms, the technology has the potential to be made available as a service to help domain experts accelerate and scale their learning process. Our research continues to push the boundaries of AI in the health domain. Ultimately, this technology may lead to a source of effective behavioral change interventions to assist health professionals, researchers, and policymakers in helping people lead healthier lives. Tailoring behaviour change interventions to specific groups of people, with certain characteristics, is associated with improved health practices and outcomes, both physical and psychological. This should increase average life spans of these groups of people and may also lead to more general societal benefits. For example, obesity, antimicrobial resistance, and hospital-acquired infections can be mitigated by healthier eating, appropriate antibiotic prescribing, and improved hygiene behaviours, respectively. Thus, advancing the understanding of human behaviour change, at both individual and population levels, can potentially have a wide-ranging impact on shaping the course of human civilization in the coming years. Unsupervised Information Extraction from Behaviour Change Literature. Debasis Ganguly, Léa A. Deleris, Pol Mac Aonghusa, Alison J. Wright, Ailbhe N. Finnerty, Emma Norris, Marta M. Marques, Susan Michie. IOS Press Vol 247 Pages 680 – 684, 2018 The Human Behaviour-Change Project: harnessing the power of artificial intelligence and machine learning for evidence synthesis and interpretation. Susan Michie, James Thomas, Marie Johnston, Pol Mac Aonghusa, John Shawe-Taylor, Michael P. Kelly, Léa A. Deleris, Ailbhe N. Finnerty, Marta M. Marques, Emma Norris, Alison O’Mara-Eves and Robert West. Implementation Science 12:121, 2017 IBM’s collaboration with The Michael J. Fox Foundation aims to better understand Parkinson's disease (PD), a chronic, degenerative neurological disorder. Grip strength is a useful metric in a surprisingly broad set of health issues. It has been associated with the effectiveness of medication in individuals with Parkinson’s disease, the degree of cognitive function in schizophrenics, the state of an individual’s cardiovascular health, and all-cause mortality in geriatrics. At IBM Research, one of our ongoing challenges […]

Yet there is a growing shortage of mental health professionals to adequately treat this need. By 2025, it’s estimated that demand for psychiatrists may outstrip supply by up to 15,600 psychiatrists (2). To help clinicians with limited resources support the growing number of patients who seek treatment, the research field of Computational Psychiatry applies data and metrics-driven approaches to psychiatry to study thought, emotion, and behavior. In January 2017, IBM made the bold statement that within five years, health professionals could apply AI to better understand how words and speech paint a clear window into our mental health. Almost two years later, we’re already seeing promising early results. Since then, the work and research we’ve done has solidified our position: individualized data—from speech to word choice to written text and physiological indicators—coupled with AI could be the key to helping health professionals better understand our own minds. Over the past year, teams from IBM Research have collaborated with clinicians to publish the following research in this space, all of which demonstrates the potential of AI and speech to help inform professionals and help them paint a more detailed picture of what’s happening within our minds. We envision a future where these technologies can be put into the hands of mental health professionals and ultimately enable them to do their jobs more intelligently, with greater confidence, and with the ability to effectively treat a growing volume of patients with the right data at their fingertips. While this is great progress, this is still just the tip of the iceberg. We’re continuing to refine and build out these techniques further, and expand their use to help clinicians get an even broader view of what could be happening within an individual’s brain when it comes to mental health and neurological disorders. Hopefully, health professionals will soon be able to frequently use speech to tap into the power of AI and make more informed diagnoses. IBM’s collaboration with The Michael J. Fox Foundation aims to better understand Parkinson's disease (PD), a chronic, degenerative neurological disorder. Grip strength is a useful metric in a surprisingly broad set of health issues. It has been associated with the effectiveness of medication in individuals with Parkinson’s disease, the degree of cognitive function in schizophrenics, the state of an individual’s cardiovascular health, and all-cause mortality in geriatrics. At IBM Research, one of our ongoing challenges […] Throughout life, many of us develop unhealthy habits that may feel nearly impossible to change. To quit smoking, reduce alcohol consumption, eat a healthier diet, or become more physically active requires effort and the right state of mind. A team of behavioural scientists at University College London (UCL)  and researchers at IBM Research-Ireland are looking […]

Today IBM introduced a comprehensive new set of trust and transparency capabilities for AI on the IBM Cloud. This software service brings to market new technologies developed by IBM Research together with Watson’s engineering team. The capabilities address the principles of explainability, fairness, and lineage in AI services, and their release is an important step towards developing trusted AI services. The services can be used with a wide variety of machine learning frameworks and AI build environments, including Watson, Tensorflow, SparkML, AWS SageMaker, and AzureML. The new services incorporate several IBM Research innovations, including checkers to detect bias in training data and models; tools to pinpoint the source of bias at various stages of the AI pipeline and to suggest mitigation strategies; and end-to-end lineage management to track the complete development of AI. These components are an initial set of capabilities that can help engineer trust into AI systems and enable AI solutions that inspire confidence. Our bias checkers consider both individual and group discrimination. Individual discrimination is when a person in an advantaged group and a person in a disadvantaged group receive different decisions when all other attributes are the same. Group discrimination is when advantaged groups and disadvantaged groups receive differing decisions on average. Disparate impact is a commonly understood measure of group discrimination. Our data bias checker analyzes the training data to find disparate impact given user-specified protected attributes. Similarly, the group model bias checker considers model predictions instead of labels in the training data. The individual model bias checker systematically generates extensive test cases to check for the presence of individually biased decisions. For easy comprehension by data scientists, the checkers present results via natural language explanations and also output relevant instances illustrating the bias that is found. In addition, the data bias checker can identify the part of the dataset that is the source of the unfairness. For example, if the data bias checker finds there to be bias against black people in a home loan dataset, the source of bias analysis might further determine that the bias was specifically against black women of a certain age. A data scientist can then use this information to supplement this part of the dataset appropriately by, for example, gathering more data. IBM’s new capabilities detect bias in AI models, capturing potentially unfair outcomes and automatically recommending ways to mitigate detected bias. In many cases, regulations like GDPR mandate that businesses maintain complete records on provenance, or lineage, of their AI platforms and components. Our new service meets this need by tracking the complete development of an AI system: data acquisition, pre-processing, model training, sharing, deployment, and retraining. The system stores information about assets (data, model, code), events affecting these assets (preprocessing, training, etc.), and entities involved in these events. At each step, the system also manages the core metadata. The system makes it easy to track the accuracy, performance, and fairness of AI applications and recall it for regulatory, compliance, or customer service reasons. It also provides explanations for an application’s decisions, including which versions of the model and data were used to make the decisions. IBM’s new Trust and Transparency capabilities can provide an explanation of why a recommendation was made. Explanations show which factors weighted the decision in one direction vs. another, the confidence in the recommendation, and the factors behind that confidence. IBM Research has developed a comprehensive strategy that addresses multiple dimensions of trust in AI solutions. We will continue to push innovations not just for checking and explaining bias, but also for debiasing or mitigating bias in data and models. We are also working to add more temporal query classes to the lineage management system to increase its applicability. Efficiently Processing Workflow Provenance Queries on SPARK Rajmohan C, Pranay Lohia, Himanshu Gupta, Siddhartha Brahma, Mauricio Hernandez, Sameep Mehta Provenance in Context of Hadoop as a Service (HaaS) – State of the Art and Research Directions Himanshu Gupta, Sameep Mehta, Sandeep Hans, Bapi Chatterjee, Pranay Lohia, Rajmohan C Researchers from MIT and IBM propose an efficient and effective method for certifying attack resistance of convolutional neural networks to given input data.