Artificial intelligence (AI) and machine learning (ML) are constantly in the headlines, from advances in access to ethical concerns about human labor and its impact on specific groups. For computer science professionals like Dr. Christelle Scharff, the emphasis is on current limitations rather than what AI can or does, especially when it comes to incomplete datasets that show the biases of AI systems. is placed.
Mr. Schaaf has been Professor of Computer Science at Pace University for 22 years and is currently Associate Dean of the Seidenberg School of Computer Science and Information Systems. When she began her research in AI as her PhD student, she described her own research as mostly “theoretical” and working on theorem-proving problems. . She was interested in studying how computers could understand mathematical and logical concepts such as deduction rules, equivalence, commutativity, and associativity. As AI technology has developed, so has her interest in machine learning and deep learning, and so has her interest in mitigating potential biases in AI. Schaaf has been focusing her research on Africa since 2009, when she received a grant to work on entrepreneurship and mobile app development in Senegal. Since then, her AI research opportunities in Africa have increased, and Schaaf and her students now continue to explore ways to ensure AI catches up with increasing global inclusivity.
As AI technology has developed, so has her interest in machine learning and deep learning, as well as mitigating potential biases in AI.
Her two recent projects with PhD students Karimuncia and Krishna Bhattuara center around African fashion. The first is expanding the reach of a popular fashion dataset called Fashion MNIST. Datasets are the pillars of the AI movement, and their creation and use require safeguards.
Fashion MNIST can identify specific garments, but fashion items that fall outside its very limited descriptions (which mostly fit into Western terminology and trends) can easily be misclassified. “If you ask this dataset to recognize a saree, it will probably say it’s a dress,” says Schaaf, for example. She explains that, like Sally, the dataset does not know how to identify specific African fashion items for her. “Because I worked in Africa as a Fulbright scholar, this project focused on African fashion, and a graduate student from Senegal was also involved.” was to create a dataset that recognizes two pieces of clothing. Bobo and red soybeans. To underscore the importance of incorporating a broader, more global language into these AI models, Schaaf explains: “If you go to a tailor and say, ‘I want a dress,’ he doesn’t know where to start.”
“Another step in any AI-related project is to ask subject matter experts.” —Schaaf
Another African fashion project that Schaaf and her team worked on is perhaps more tangible to people outside the AI community. They worked to recreate patterns popular in Africa. wax– Wax shiny colorful geometric patterns.
The team collected a dataset of about 5000 free wax patterns and created new satisfying patterns generated by AI. From there, the team printed selected patterns and partnered with local artisans in Senegal to create fashion items such as bags. A dataset was constructed to generate different patterns. Schaaf estimates that more than 10,000 different images are needed to get a good sample. If the dataset is mostly blue patterns, the generated patterns will be pinned predominantly to blue. Alternatively, if you don’t have enough flower images, you’ll need to add those images to your dataset to get them. The very nature of exclusion changes what AI can produce, demonstrating the need to expand these datasets to reflect the world as it is, not just what has been input.
Her students work hard to build that dataset and corresponding models to generate interesting wax patterns.
“My biggest concern is diversity bias. But I think the discussion is more open now. At least everyone is aware of this problem. Having policies, tools, processes and practices for “
— Shaaf
For those worried about AI stealing their jobs? “Another step in any AI-related project is to ask subject matter experts,” says Schaaf. She explains that once these patterns are created, they need to be reviewed by fashion experts to understand what’s working and what’s not. Computers can create patterns, but they can’t (currently) classify what is fashionable for everyday wear, what is artistic, or what aesthetic category it belongs to.
Schaaf is excited about where AI is headed and how ubiquitous it will become. Her primary concern is exactly what her job does: balancing datasets to make them representative and diverse. “My biggest concern is diversity bias. Having the policies, tools, processes and practices to make it happen.”
