Saška Mojsilović wants to channel AI for good. She may also make you rethink sour cabbage
A data scientist and an accomplished food writer, Mojsilović has focused her creative energy on building trustworthy, beneficial technologies and finding connection through food.
A data scientist and an accomplished food writer, Mojsilović has focused her creative energy on building trustworthy, beneficial technologies and finding connection through food.
In a field of big personalities, Saška Mojsilović knows how to look the part. On a cold February day, you might find her by the windows in the IBM Research library sporting a sweater that reads ROCK & ROLL, words that speak to her playful but unwavering leadership style.
“When Saška shows an interest in something new, people take note,” a longtime colleague, Mahmoud Naghshineh, noted. “And once she believes in it, good luck changing her mind.”
Currently head of Trustworthy Machine Intelligence, Mojsilović helped convince her bosses to develop the tech industry’s first tools for combating algorithmic bias, reinforcing the idea that creators of technology should hold themselves accountable for the safety of their products.
As a student in the former Yugoslavia, Mojsilović anticipated the rise of big data, applying machine learning algorithms to disease diagnostics, and later, to recommendation engines, workforce analytics, and IT operations. As a reward for laying the foundation for IBM’s data science business, she was named an IBM Fellow — Research’s equivalent of a rock star.
More recently, Mojsilović has branched into AI for drug discovery while continuing to advocate for the ethical use of AI. “It should be the first lens we put on when we think about creating technology,” she said. “What are the benefits and what are the problems? Can the technology be used to create harm or unfairly benefit some groups over others?”
When she’s not writing code or debating AI’s existential questions, she is running experiments in her kitchen for her blog, Three Little Halves, a nominee for a James Beard Journalism Award. Mojsilović may be one of the few people on the planet who can write as passionately about sour cabbage as they can about statistical distributions. In photos, essays, and recipes for delicacies like caramelized sauerkraut and bite-sized, jam-filled Serbian holiday cookies, Mojsilović explores family traditions, memory, and loss. The Yugoslavia of her youth, torn into six countries after years of ethnic war, figures prominently.
When Mojsilović was seven, her father gave her his old Leica knock-off and a light meter. She’d practice taking photos in a park near her home in Belgrade before loading the camera with 35 mm film. In the 1970s, buying and developing film was expensive: “You didn’t just go around shooting pictures like crazy,” she said.
Not only were photos precious, their contents — the scene, colors, and composition — were fixed once they were printed. The rise of the personal computer would change that. For Mojsilović, the command line of the Commodore 64 her parents bought for her as a teenager opened a portal to a new world. She tired quickly of playing videogames on it and turned to learning how to code and manipulate images. Computers appealed to her creative side and could produce as many images, with as many permutations, as she desired.
When she compiled her first program, she gasped at the result. The white outline of a house appeared on the black backdrop of the computer monitor. A working edge detector. “It was the biggest excitement of my life,” she said. She was in college, and that sense of delight never really faded.
Looking for new projects, a professor connected her to a doctor who was using machine learning to analyze ultrasound images from Serbia’s lone digital machine. Today, cancer-detecting algorithms are routinely trained on tens of thousands of images. But at the time, Mojsilović was grateful to load 20 scans on a floppy disk to analyze. “It was so precious to me,” she said. “I got to know every blood vessel. They were imprinted in my head.”
She earned a PhD in image processing at the University of Belgrade, as Yugoslavia was disintegrating into violence. When she graduated in 1997, the economy was in tatters and there were few jobs to be found. So, she fell back on tennis, a sport she had played at a national level as a teenager, once even facing former world No. 1 Monica Seles in a match. As a tennis coach to Belgrade’s elite, she made more than her mother and father (a lawyer and civil engineer, respectively) combined. But after four months, she decided she’d be happier teaching at the university.
The story might have ended there were it not for a chance encounter with a Belgrade scientist at a conference in Switzerland. The researcher invited Mojsilović to join her at Bell Labs, and though very few Serbians were being welcomed abroad at the time, Mojsilović made it to the United States. She moved to IBM Research after two years at Bell Labs, where she met her husband — known as Dr. V in her blog, who with her daughter, Miss Pain, comprise her other halves.
When Mojsilović joined IBM in 2000, Amazon was still mainly a bookstore, and recommendations for everything from movies to plumbers came mostly by word-of-mouth. The rise of the internet not only made everyone a critic, it democratized access to information, which conveniently was machine-readable. Anyone with technical chops could mine it for insights.
Mojsilović moved to what was then IBM’s math department, where she built models that combined a burgeoning flood of digital information with in-house data to answer practical questions: Where in the country should IBM place individual salespeople to best use their talents? What products should it pitch a particular client in a softening economy? When was the right time to outsource?
The job involved martialing algorithms to find patterns in a firehose of data so IBM and its enterprise clients could make evidence-based decisions, work that today is known as data science. She published several innovative papers during this time, introducing a new statistical method to mine big datasets and a methodology for managing outsourcing projects. The work was profitable and led to her being named an IBM Fellow in 2014 — the company’s highest honor. An entirely new business unit in data analytics later sprung from this work.
Big data was still a buzz word when Mojsilović moved to IBM Research’s AI team to focus on AI trustworthiness and fairness. With enough data and compute, machine learning models held promise to improve or automate many applications. But they had an Achilles heel: their decisions were often impossible to trace back to the source data to ensure they were accurate and fair.
Mojsilović understood this early on. As a scientist, she was used to diving deep into her data to understand its limits before writing any code. She knew that if the data were bad or incomplete, an AI model would produce incorrect or biased results. Yet she was surprised to learn that the police, courts, and other institutions were already using these tools to make high-stakes decisions without understanding their vulnerabilities.
In 2017, she and her team urged IBM management to develop bias-mitigation software. Initially, their plea was met with skepticism, she said. But they persisted, and a year later, IBM became the first tech company to launch a free library of de-biasing algorithms, the AI Fairness 360 (AIF360) toolkit, and to incorporate bias mitigation and explainability into its own products.
“The idea was to help people understand bias and learn how to check for it – to basically adopt good data-science practices,” she said.
A pair of complementary toolkits, AI Explainability 360 and FactSheets for AI services, quickly followed. FactSheets allowed ordinary people to gauge an AI system’s credibility the same way nutrition labels or energy ratings let you judge the healthiness of a granola bar or the efficiency of a boiler. They told consumers where the model’s training data came from, which algorithms were used, and whether any bias checking or mitigation had been done. Today it is central to IBM’s AI Governance solutions.
The tools emerged from IBM’s Science for Social Good program, an earlier initiative launched by Mojsilović and her colleague, Kush Varshney. The main lesson they took away was that to tackle systemic problems, you had to create solutions that were easy to use and share and could be generalized across organizations and applications.
The experiment convinced Mojsilović that open-source software was the best way to quickly disseminate bias-mitigation tools and best practices. Today, AIF360’s GitHub repository has more than 700 stars, and numerous IBM customers and developers have incorporated it into their applications, she said. Engineering schools also call upon her often to discuss the work.
Her message is consistent. “Papers are really no good if you don’t put them into practice,” she told a class of NYU students over video last fall.
If putting ideas into practice has become a Mojsilović trademark, so has including a social good angle in all that she does. “It’s in her blood,” said IBM’s Payel Das, a frequent collaborator. “She always tells us, don’t just work toward one more paper or product but think about its impact. She’s good at seeing the big picture.”
Considering a new technology from every angle doesn’t come naturally, Mojsilović insists. It takes work. “We can’t make good technologies until we as humans do good,” she explained. “Our technologies are the reflections of us – how we think, what we believe in, what we care about. Cultivating that is tremendously important.”
Her alter ego, Queen Sashy, puts it another way in an essay about Mojsilović's prize-winning recipe for vanilice cookies that has been passed down over generations. “If we all carried spatulas and rolling pins instead of something else, if we all told stories, and took photos, and cultivated a little bit more love, this world would be a much better place.”