Caltech Researchers Pioneering Ethical AI and Computer Vision

At Caltech, researchers are at the forefront of developing artificial intelligence (AI) technologies, particularly in the field of computer vision. This branch of machine learning focuses on teaching computers to interpret and understand visual data, such as images and videos. However, the development of these technologies is not just about technical prowess; it also involves addressing ethical concerns to ensure that AI is fair and just for all users.

Pietro Perona, Caltech's Allen E. Puckett Professor of Electrical Engineering, is an AI pioneer in computer vision. His team has been working on algorithms that enable machines to recognize objects with minimal human supervision. One of the key challenges in this field is the collection and use of large datasets, which can introduce biases if not handled carefully.

"We have to collect very large datasets," Perona explains. "This step is sensitive. Do you own the data? Are you asking for permission to use it? If you can download the data from the internet, is it reasonable that you use it? Do the data contain biases that may affect the algorithm?"

For instance, training a computer to recognize birds using only images taken on bright summer days can lead to an AI system that performs poorly at night. Ethical questions become even more critical when AI is used to make decisions that impact people's lives, such as filtering job applications or predicting criminal behavior.

Perona and his collaborators, including Colin Camerer, the Robert Kirby Professor of Behavioral Economics, have developed a method to measure algorithmic bias in vision language models. These models can analyze both images and text, and the team used AI to generate a dataset of realistic human face images that were systematically varied across age, gender, race, facial expression, lighting, and pose. They also created a dataset of text prompts based on psychological research.

The researchers fed these images and text prompts into a popular vision language model called CLIP and analyzed how the model represented the data. They found that the model contained biases, particularly in how it perceived Black women. For example, frowning Black women were perceived as the least competent, while smiling Black women were perceived as the most competent. This study provides a benchmark for AI engineers and researchers to test and improve their own models.

Perona emphasizes the importance of responsible AI development. "Engineers can provide numbers and statistics about our AI models, but it’s up to society, through the law and elected leaders, to figure out a consensus on what is fair and ethical in different contexts," he says. Caltech aims to teach AI principles to all students, including the ethics of responsible AI.

Yisong Yue, a professor of computing and mathematical sciences at Caltech, agrees that computer scientists should consider the ethical implications of their work. However, he notes that much of the time, researchers are working on early-stage prototypes that need refinement before practical deployment. "We typically design tools and then partner with industry in deploying them," Yue says. "When we see something beginning to work, that’s when we think about the more practical implications."

AI is also being used to combat misinformation. Michael Alvarez, the Flintridge Foundation Professor of Political and Computational Social Science, uses AI to help people identify and resist online falsehoods. His research involves using generative AI to create prebunking warning labels, which can help people build mental resistance to misinformation.

Alvarez is also co-director of the Linde Center for Science, Society, and Policy, which brings together researchers, policy stakeholders, and industry professionals to discuss AI's social, political, and economic impacts. The center aims to understand how AI technologies are driving change and to inform policy on pressing societal issues.

"One of our goals is to try to understand, as best we can, how all of these new artificial intelligence technologies are driving this broad area of social, political, and economic change," Alvarez says. The center organizes forums to discuss these topics and provide scientific expertise to inform policy.

Caltech's commitment to responsible AI development is evident in its interdisciplinary approach, combining technical expertise with ethical considerations to ensure that AI technologies benefit society as a whole.