Artificial intelligence is not only the next evolution of computing, it is also helping to define the future of human knowledge and the possibilities of advanced cognition.
This month, we are highlighting the work of four AI researchers at IBM who are pushing the frontiers of the technology. Their efforts extend from work process automation to the design of ever more intelligent chatbots to the discovery of new, more effective antibiotics. All four of these researchers are women—a constituency that has helped lead IBM Research in the crucial task of removing or mitigating bias from AI algorithms—a key for fairness and gender equity.
Teaching Chatbots from their Stumbles
For Inbal Ronen, mistakes are opportunities. Ronen, a 16-year veteran at IBM Research in Haifa, Israel, focuses on the stumbles of chatbots. Every time one of them falters—failing to understand a question or botching an answer—Ronnen sees a teaching opportunity. As she sees it, her job is to advance this educational process for AI.
IBM’s customers, Ronen says, use Watson Assistant to improve service. Customers are able to get quick answers without waiting on help lines, and human agents are able to devote more time to more complex questions. She zeros in on incidents where bots get confused and hand a query over to a human. Sometimes, she and her team study the human response, and then use that to instruct the bot. The more efficient method, however, is to engineer the system itself to learn from the human, and adjust automatically. “In that sense,” she says, “the human is teaching the bot.”
Ronen studied math and computer science in Israel, and got her master’s degree in computer science in Jerusalem. She stayed there early in her career, working at several startups. Her specialty was the exploding field of social search and social network analysis.
In Jerusalem, she met her husband, who is also a technologist and a former IBMer. They have three children. “I’m a full-time working mom,” Ronen says. It’s a dual job that involves education of human beings as well as machines.
A Scientific Approach to AI Discovery
How do you increase the chance of scientific success? Payel Das and her team at the T.J. Watson Research Center in Yorktown Heights, N.Y., are turning to physics to help resolve that problem. “We are developing machine learning algorithms that can combine learning from not only data, but also from physics principles, in order to design new materials and drugs,” says Payel, a Research Staff Scientist and Manager of Trusting AI research. “When we combine machine learning, scientific knowledge and a set of rules, the success rate of new scientific discovery can go up 100-fold.”
Using this approach, Das and her team developed an AI algorithm that can find novel antimicrobial peptides that could eventually be used to develop new antibiotic drugs, a discovery they hope to soon publish in a major scientific journal.
The infusion of science will help ensure machine learning is robust, interpretable, fair and creative. “We don’t just want predictions from AI, we want to see if a model can explain why something is, or isn’t, going to work,” adds Payel, who has published more than 40 peer-reviewed articles and is an adjunct associate professor in Columbia University’s Department of Applied Physics and Applied Mathematics (APAM).
Payel faced many obstacles on her path to IBM Research. Growing up in Kolkata—the capital of the Indian state of West Bengal—the idea of girls pursuing any career, much less one in math or science, was not widely accepted. “My mother earned a bachelor’s degree in history in the 1970s, but could not pursue her studies further because her family was not very supportive,” she says. “That motivated me because, in a sense, she had to compromise her career because of her family.” Fortunately, Payel had no shortage of support from her immediate family, in particular her parents and an uncle who taught chemistry.
After receiving her bachelor’s and master’s degrees in chemistry in India, Payel moved to the U.S. in 2002 to pursue a Ph.D. in theoretical chemistry at Rice University in Houston. Her interest in seeing rapid, more tangible results from research led her to IBM Research in 2007.
Payel, who is married to an experimental chemist and has an 11-year-old daughter and four-year-old son, finds motivation in the challenges she faces as a woman engaged in a STEM career. “If a young woman is passionate about pursuing a particular area,” she says, “I would advise her to go for it no matter the obstacles or what the statistics say.”
Making AI More Human
As AI becomes more prominent, so do fears that the technology will put people out of work. Yunyao Li wants to put much of that fear to rest. She and her team at IBM Research – Almaden are investigating ways to ensure humans remain a critical part of AI training and decision making.
“There are a lot of things that data alone cannot tell you or that are more easily learned by asking someone,” says Yunyao, a Principal Research Staff Member and Senior Research Manager for Scalable Knowledge Intelligence. “That’s the beauty of [having a] human in the loop.”
IBM’s human-in-the-loop research investigates how best to combine human and machine intelligence to train, tune and test AI models. Yunyao is leading a group investigating how to apply this approach to help AI better interact with people through natural language.
The HEIDL (Human-in-the-loop linguistic Expressions wIth Deep Learning) model they introduced last year proposes to bring expert humans into the AI loop twice: first to label training data, then to analyze and improve AI models. In their experiment they described using HEIDL to improve AI’s ability to interpret the dense legal language found in contracts.
Yunyao and her colleagues are working to advance last year’s research by better automating data labeling and improving HEIDL’s ability to interpret words not included in training dictionaries. Some of her other Natural Language Processing (NLP) research is aimed at helping train expansive AI systems using unstructured data, “a service that hasn’t been available to enterprises in a scalable manner,” she says. “I focus my work on NLP because language is the most important medium for human to share information and knowledge. NLP essentially helps machines to read and write, and thus learn to learn and share information and knowledge with people.”
Growing up in the 1980s in Jinsha, a small town in southwest China, Yunyao had little exposure to computers. “Due to the poor economic situation at the time, I traveled outside our hometown only a couple of times before I went to college,” she says. One of her favorites books growing up was Jules Verne’s Around the World in Eighty Days. “The book’s fascinating stories of technology and travel inspired me to travel, explore unknown places and learn about different technologies and culture,” she says.
Yunyao enrolled in Tsinghua University in Beijing, where she ranked at the top of her class and received a dual undergraduate degree in automation and economics. Her interest in technology next took her to the University of Michigan, where she earned master’s degrees in information science as well as computer science and engineering. By 2007, she had likewise earned her Ph.D. in computer science from Michigan.
Positive experiences with mentors in school and as a young professional have inspired Yunyao to take on that role for a new generation of women computer scientists. “It was very challenging to me when I moved from China to Michigan,” she says. “Fortunately, as a student I found a wonderful mentor—Mary Fernandez, a researcher at AT&T Research. Like myself, part of her family was living oversea at the time, and she was in a long-distance relationship with her husband for a few years, so we could relate to one another.” Yunyao’s husband, Huahai Yang, moved from Michigan to join the faculty at the State University of New York – Albany shortly before they got married and were in a long-distance relationship for a few years.
Yunyao has benefitted from several mentors at IBM, as well, including Almaden researcher Rajasekar Krishnamurthy, former IBM Fellow Shivakumar Vaithyanathan and Laura Haas, who retired from IBM Research in 2017 after 36 years. “Now, I want to share my experience with other people, and help give young researchers some visibility into their own future,” she says.
Focusing AI on Human Trafficking
Prerna Agarwal wants to make one thing clear. “I owe my career to my mother,” she says. “She left her job as a teacher and sacrificed to raise us.” Backed by her supportive family, Agarwal went to university in New Delhi and later received her master’s in computer science from the Indraprastha Institute of Information Technology (IIT Dehli). In 2017, she joined IBM Research in New Delhi. She specializes in AI.
Now she uses AI to help children who are far less fortunate: the estimated 1 million Indian teenagers who are victims of human trafficking. Thousands of them are rescued every year, but they’ve suffered searing trauma–physical, mental and sexual–and need counseling. The trouble is that there are not nearly enough trained counselors to help them.
This is where Agarwal’s AI can help. Working with a non-profit called EmancipAction, she is developing a system to analyze resumes, questionnaires and video interviews to pinpoint the most promising candidates to be trained as counselors for trafficking victims. The AI, she says, scouts for bias and gender awareness, and analyzes video and speech for signs of emotional intelligence. The system will grow more robust, she says, as it processes the feedback and adjusts its predictions.
In addition to her work for social good, Agarwal develops AI systems for business processes. One focus is to analyze work processes, scouting out areas of inefficiency, so-called hot spots. She and her team zero in on these bottlenecks, studying the various tasks involved. They build systems to speed up the work, providing decision recommendations. At the same time, they identify steps in the process that can be automated.
Before Agarwal and her team can program software to handle a job, they need to dissect the task into its base components and identify every decision point. Building even the most sophisticated AI, after all, often means asking the simple questions that most humans never bother to ask. “We have to identify who are the actors involved,” she says “There’s a finite set of them. What are the steps that they’re taking, and how complicated are they?” It’s through this process, she hopes, that she will contribute to AI systems that give back to society.
This post is presented by The Watson Women’s Network, a community of technical staff, primarily based at the T.J. Watson Research Center, that seeks to encourage a workplace environment that advances the professional effectiveness, individual growth, recognition, and advancement of all women at IBM Research. The WWN partners with senior management, human resources and other diversity network groups to promote programs in mentoring, networking, diversity, knowledge sharing and recruiting.