Bosch Center for Artificial Intelligence, Carnegie Mellon Partner To Accelerate AI Research

SCS Professor Zico Kolter Will Join Bosch as Chief Scientist of AI, Remain on Faculty

Bosch in North America today announced the launch of the Bosch Center for Artificial Intelligence (BCAI) Research Lab in Pittsburgh, which will be the BCAI's fourth location. The lab will conduct advanced research in artificial intelligence (AI) technologies. The new location is the next step in BCAI's mission to partner with leading institutions around the world to jointly accelerate AI research. Bosch plans to build a team of up to 20 AI experts at Bosch's Pittsburgh Technology Center by the end of 2019. In addition, Bosch will provide more than $8 million to sponsor research at Carnegie Mellon University through 2023. This collaboration represents another important milestone in the longstanding and mutually beneficial partnership between Bosch and the university. Having established a presence in Pittsburgh in 1999, Bosch has played an important role in fostering the development of Pittsburgh's flourishing tech community. The launch of the BCAI Research Lab in Pittsburgh is highlighted by CMU's Zico Kolter joining Bosch as chief scientist in AI. Kolter, an assistant professor in CMU's School of Computer Science (SCS), will direct the multi-million dollar research projects at CMU, in addition to contributing to Bosch's global R&D efforts as a principal member of the BCAI. Kolter, a leading expert in AI research, develops methods that make machine learning more robust, interpretable and modular. He also has worked on applications for smart energy and sustainability solutions. Some of his recent work looks specifically at machine learning methods that are provably resistant to so-called "adversarial attacks" and methods for incorporating optimization procedures as modules within the loop of more complex deep learning systems. Significantly, while working for Bosch, he will continue to teach and perform research as a faculty member at SCS, which U.S. News and World Report this spring rated the No. 1 U.S. graduate school for artificial intelligence, as well as a top-ranked school for computer science overall. "We're excited to establish BCAI Research in Pittsburgh with Zico Kolter as part of the long-term collaboration between CMU and Bosch," said Christoph Peylo, global head of BCAI. "CMU, with its tradition as one of the leading institutions in AI research, is an important pillar in BCAI's mission to develop safe, robust and secure AI for Bosch products and services." With a widening gap between industry need and available talent, Bosch is partnering with leading academic organizations such as CMU to develop future professionals' skills in the dynamic field of AI. The collaboration will also create new opportunities to further AI technology that contributes to advances in Bosch's four business sectors — mobility solutions, consumer goods, industrial technology, and energy and building technology. Under this partnership, Bosch will expand its AI capabilities while simultaneously training the next generation of AI experts. "We at Carnegie Mellon are excited to be working with Bosch to find new ways to use artificial intelligence to improve lives and to develop the next generation of AI scientists," said Andrew Moore, dean of CMU's School of Computer Science. "Bosch will benefit greatly from Zico Kolter's insights, while Zico continues his essential research and teaching here at Carnegie Mellon." CMU has been a leader in AI since the field emerged 60 years ago, pioneering advances in self-driving cars, facial recognition and natural language processing. Last year, SCS expanded its efforts through the CMU AI initiative and this year launched the first U.S. undergraduate degree in AI. BCAI was founded in early 2017 to deploy cutting-edge AI technologies across Bosch products and services. Using data from Bosch domains, BCAI creates differentiating technology by solving challenges in high-potential lead applications. It scales research results and generates value through deployment of machine learning in products, processes, and services in areas such as manufacturing, engineering, supply chain management and intelligent services. BCAI Pittsburgh is the most recent addition to BCAI's locations around the globe — Renningen, Germany; Sunnyvale, California; and Bangalore, India. Since its inception, BCAI has sought out innovative partnerships with international universities as well as partners from politics, business and science to solve today's most challenging AI problems. BCAI is actively looking for opportunities to expand its research network further and collaborate with industry thought leaders. During the Robotics: Science and Systems (RSS) conference hosted at CMU June 26–30, Bosch and its partners will celebrate the launch of the BCAI research lab, as well as the partnership with CMU, with robotics researchers from around the globe. This invitation-only event will be held June 27 at 8 p.m. at Phipps Conservatory near CMU's campus in Pittsburgh, following the Bosch-sponsored RSS Pittsburgh Robotics Showcase. The Robotics Showcase will be held in the Atrium of CMU's Newell-Simon Hall and is open to all conference attendees.

Computational Method Puts Finer Point on Multispecies Genomic Comparisons

Probabilistic Model Could Provide Insights Into What Makes a Human a Human

A new computational tool will potentially help geneticists to better understand what makes a human a human, or how to differentiate species in general, by providing more detailed comparative information about genome function. In a report published online today by the journal Cell Systems, researchers led by Jian Ma, associate professor of computational biology at Carnegie Mellon University, describe a new model for performing comparative analyses of genome function across multiple species. Such analysis may provide insights into not only evolution, but also human disease. The research team, including scientists from the University of Virginia, Florida State University and the University of Connecticut, developed the Phylogenetic Hidden Markov Gaussian Processes model, or Phylo-HMGP, to analyze functional genomic data. They used the model to analyze a new dataset for DNA replication timing across five primate species, including human. Genetic differences in protein-coding genes alone cannot account for the dramatic variation between species, so scientists increasingly focus on differences in gene regulation — mechanisms that control how and to what degree genes are activated. "The differences among primate species may be mostly in the noncoding regions of the genome, the regulatory elements, not the genes themselves," Ma explained. High-throughput technologies produce a large amount of functional genomic data, which should help scientists better understand how genomes evolved. Ma said Phylo-HMGP addresses what might be called the "Starbucks problem" in these multispecies analyses. Just as coffee vendors tend to sell drinks in small, medium and large sizes, analysis tools typically characterize functional genomic data as low, medium or high. "With Phylo-HMGP, we can look at each functional genomic value as a continuous signal — showing the actual activity level rather than just a rough level estimate," said Yang Yang, a Ph.D. student in CMU's Computational Biology Department and first author of the study. "In this way, we're able to fully utilize the data that have been gathered." The researchers applied the model to an analysis of DNA replication timing, the order in which segments of DNA are replicated, which can vary from species to species. They did so for a dataset including humans, chimpanzees, orangutans, gibbons and green monkeys that was generated in collaboration with David M. Gilbert of Florida State University and Rachel J. O'Neill of the University of Connecticut. "We demonstrated that we could use Phylo-HMGP to discover genomic regions with distinct evolutionary patterns of replication timing," Ma said. Their research provides a framework for applying the model to reveal genomic regions with functions that are similar across species and those that are varied, or dynamic, between species. Analyses of dynamic regions in functional genomic datasets not only can improve understanding of evolution, but also may have implications for certain types of species-specific diseases, he added. Other research team members include Yang Zhang, a research associate in the Computational Biology Department; Quanquan Gu of the University of Virginia; Takayo Sasaki of Florida State; and Julianna Crivello of the University of Connecticut. The National Institutes of Health and the National Science Foundation supported this research. A copy of the research paper is available on the Cell Systems website.

Carnegie Mellon Names Rosenfeld Head of Machine Learning Department

Carnegie Mellon University's School of Computer Science has named Roni Rosenfeld, an expert in epidemiological forecasting and spoken dialogue technologies, head of its Machine Learning Department. He succeeds Manuela Veloso, who is taking a leave of absence. Rosenfeld will take the department reins on July 1. Rosenfeld, a professor in both the Language Technologies Institute (LTI) and the Machine Learning Department (MLD), joined Carnegie Mellon in 1994. "Machine learning is incredibly important in today's world, and there are few people I trust as much as Roni Rosenfeld to lead our efforts in this area," said Andrew Moore, dean of the School of Computer Science. "Roni's work in forecasting disease and his dedication to machine learning for good make him an excellent choice for this role, especially as we strive to make CMU the hub for artificial intelligence that changes the world for the better. I'm thrilled that he will run our Machine Learning Department." As head of CMU's Delphi Research Group, Rosenfeld aims to make forecasting disease as universally accepted and useful as forecasting the weather. In particular, his group has forecasted the flu using two methods. The first relies on artificial intelligence — specifically, machine learning — to make predictions based on past patterns and input from the Centers for Disease Control and Prevention's domestic influenza surveillance system. The other method relies on the so-called "wisdom of the crowd," basing its forecasts on the judgments of a number of volunteers who submit their own weekly predictions. Last fall, the CDC reported that these two models proved to be the most accurate among all forecasting systems it evaluated — for the third year in a row. Rosenfeld's research also focuses on finding ways to use spoken language technologies — automatic speech recognition, speech synthesis and human-machine dialog systems — to aid global socioeconomic development. One such project, Polly, uses telephone-based viral entertainment to reach low-literate people, familiarizing them with speech interfaces and then introducing them to development-related services. A caller records a message and Polly adds funny sound effects, such as changing a male's voice to a female's voice (or vice versa). The caller can then forward the message to one or more friends, who in turn can forward it along or reply to it. In Pakistan, where Polly was first launched, and later in India, the phone service linked people with recordings of job advertisements. In Guinea, Polly linked to health information recordings about the Ebola outbreak. In addition to teaching and engaging in his own research, Rosenfeld also oversees CMU's Machine Learning for Social Good fund, which provides opportunities for faculty and students to apply their expertise in data science and machine learning to initiatives that benefit the public sector. CMU uses the fund, established by a $1 million donation from the Chicago-based predictive analytics firm Uptake, to support research projects for nongovernmental organizations, nonprofits and government agencies. "More than twenty years ago, Tom Mitchell, Steve Fienberg and other faculty members in computer science and statistics had the foresight to establish the world's first academic department devoted to machine learning," Rosenfeld said. "Since then, it's become a world leader in machine learning research and education, under the able stewardship of Tom and, most recently, Manuela. I am honored to be entrusted with overseeing this incredible institution." Rosenfeld earned his bachelor's degree in mathematics and physics at Tel-Aviv University in Israel before joining the CMU community to pursue his master's and doctor's degrees in computer science. His work on speech recognition technologies earned the university's Allen Newell Award for Research Excellence in 1992, and he joined the CMU faculty in 1994. He's been a full professor in the LTI and MLD since 2005, with appointments in the Computer Science and Computational Biology departments, and the Heinz College of Information Systems and Public Policy. Rosenfeld has earned the School of Computer Science's Joel and Ruth Spira Teaching Excellence Award, and has served as associate editor of IEEE Transactions on Speech and Audio Processing. He has advised more than 20 students, taught two popular courses in language and speech and machine learning, and served on the university's Faculty Senate.

AI for Good: A Spinoff Success Story

The following story was published by Marinus Analytics, a spinoff of Carnegie Mellon University's Robotics Institute that uses artificial intelligence, machine learning, predictive modeling and geospatial analysis to combat sex trafficking. (Reprinted with permission from Marinus Analytics. Names have been changed to protect identities.)The OutcryIt started with an outcry from Allie, who was pimped two years ago by a violent trafficker who went by the name Julian. She told Detective John Patterson, "I want to get out of this because of what Julian's done to me. And he did it to a 15-year-old girl, too." Similar cases of this scope — which grew to 21 identified victims — would usually take a year. Detective Patterson built this case in about three months. He credits the importance of good experience, training and technology tools like Traffic Jam.The Violent PimpJulian was a violent pimp who required a $1,500-per-day quota for each victim and would beat any violators. He had been arrested many times in the past —for running his victims over with a car, for strangling one of his victims until she passed out, for punching and assaulting others. He also threatened to kill his victim's children if they didn't work for him.Julian recruited his victims in person and on social media apps. He broke one victim down by recruiting her to work as a stripper, and then repeatedly raping her. When she still refused to sell sex for him, he withheld food until she agreed. When she tried to escape, he used a location-tracking app on her phone to chase her down. He found her, assaulted her and put her back to work.Traffic JamDetective Patterson used Allie's testimony to begin piecing together the case. By searching victim's Facebook photos through Traffic Jam's FaceSearch, he found their ads posted across the country. "I used Traffic Jam to map out the course that Allie exactly described," he said.He was searching for Jessica, an underage victim Allie had told him about. He scrolled through Jessica's Instagram, and found the most recent pictures she posted of herself, which were more than two years old — from when she was 15. "I didn't think it would lead to anything, because it was such an old photo," he said. "But I thought I'd run it through FaceSearch just in case. I couldn't believe it when the two-year-old photo returned top matches in FaceSearch that looked just like her." The Traffic Jam trail showed that she was posting from California, but had recently posted in his city.Allie told him about another victim, Sammy. He found some year-old pictures of Sammy on her Facebook profile and uploaded them into FaceSearch, which returned top matches that looked nothing like Sammy. "I thought the matches weren't her, they just didn't look like her," Patterson said. He sanity-checked the top matches by checking the timing and location of the ads. Then, he said, "I found that one of the phone numbers in the ad was registered to her name. That made me realize that the pictures from the FaceSearch results were a correct match, but I didn't recognize her at first because she had changed her appearance so drastically." When the appearance of the victim looked completely different, FaceSearch still made a positive match in seconds.Where Are They NowBy using technology tools like Traffic Jam in conjunction with victim interviews and evidence gathered through search warrants, Detective Patterson assembled a history of money transfers from the victims to their pimp. He confirmed that many of Julian's victims were working in different states and wiring their earnings back to Julian. He determined Julian was making about $15,000 a month from two girls alone, and he had a total of 21 victims during the span of the investigation.The police department received an arrest warrant for Julian for six felonies. Today, this violent trafficker is in jail without bail, awaiting prosecution and potential life in prison for his crimes, all thanks to the tireless efforts of Detective Patterson and his team.

Carnegie Mellon Cameras Provide 24-Hour Monitoring of Industrial Sites

New Mon Valley Breathe Cams Document Sources of Visual Pollution

Carnegie Mellon University's CREATE Lab is using a new network of cameras to provide 24-hour monitoring of visible air pollution from industrial sources in the Monongahela Valley. The network is being activated today in conjunction with the United Nations' World Environment Day.The high-resolution cameras are trained on three plants in U.S. Steel's Mon Valley Works: the Clairton Plant coke works, the Edgar Thomson Plant in Braddock, and the Irvin Plant in West Mifflin. Their images will give residents and government officials a means of detecting, monitoring and documenting sources of smoke and other visible air pollution.The Mon Valley Breathe Cams will join the CREATE Lab's existing Breathe Cam network. Since 2014, the camera system has provided panoramic images to help people in the region visualize what the air over southwestern Pennsylvania looks like on different days and under different atmospheric conditions. The systems were made possible by grants from The Heinz Endowments.Randy Sargent, who directs Breathe Cam for the CREATE Lab, said the new Mon Valley cameras are being deployed in response to citizen reports to the SmellPGH app. Created by the lab with a grant from The Heinz Endowments, the app helps people share reports of unpleasant or peculiar smells with each other and with the Allegheny County Health Department."Many of the smells being reported appear to emanate from the Mon Valley, based on the report locations, prevailing winds and prevalence of pollution-trapping temperature inversions in the river valley," Sargent said. "We hope to get a better idea about what is happening by observing three of the valley's largest pollution sources."Though all three U.S. Steel plants already are monitored by the health department, cameras make it possible to detect fugitive emissions — air pollution that occurs episodically and doesn't necessarily escape through a smokestack, Sargent said. Cameras also capture the color of emissions, which provides clues about their chemical content, and show where pollution goes.The Shenango Channel, an earlier offshoot of the Breathe Cam that monitored the Shenango Inc. coke works on Neville Island, proved the value of 24-hour visual monitoring. Before DTE Energy shuttered the plant in 2016, the Shenango Channel provided evidence of fugitive emissions that resulted in citations for violating county air regulations."The beauty of these cameras is they provide visual images of the smells that many have been smelling and reporting on the SmellPGH app," said Mark Dixon, a local environmental documentarian who has had early access to the Mon Valley cameras. "It's one thing to say the region suffers from bad pollution on temperature inversion days, and quite another to see what happens when plumes of smoke hit an invisible wall and spread out over the valley."As with the existing Breathe Cam network, users can operate interactive controls to zoom in on items of interest, whether it is a hovering brown cloud or an individual smokestack. They can scan back in time to look for sources or particular incidents, and can skip back to particular dates and times of previous incidents. The tools were designed to be easy to use and learn. A short tutorial helps users and residents to capture footage that identifies and quantifies events, such as releases from smokestacks.The additional cameras, which also capture images more frequently and in higher resolution, have increased the volume of imagery processed by Breathe Cam network by a factor of 10, necessitating some technical upgrades of the system."My hope is that the cameras will make someone accountable when things go wrong," said Mary Carey, a 25-year resident of Braddock. "There should be consequences to pay when polluters don't do what they should be doing to protect people's health.""I've lived in Clairton all my life and have seen many of my classmates suffer from asthma," said Jaden McDougald, a graduating senior at Clairton High School. "I'm excited that there will be a tool that people in my community can use to learn more about what is contributing to our poor air quality and how we can fix it. It's about time Clairton got some fresh, clean air."