Jay Pujara, research assistant professor and research lead at ISI, teamed up with Lise Getoor, professor of computer science and engineering at UC Santa Cruz, and William Wang, assistant professor of computer science at UC Santa Barbara, for an award-winning research project. In 2020, the three researchers were granted around $400K in funding from Google for their groundbreaking ideas on creating probabilistic models for AI systems. Essentially, probabilistic AI models entails teaching AI how to predict the probability of an event with given information, such as using the weather forecast to predict mood, which is then used to inform restaurant recommendations.
Bringing together a strong mix of background and expertise, the three researchers decided to collaborate again for their newest project with the help of experts from Google.
While recommending systems already exist and are widely used, this model takes it to the next level by analyzing context clues to ensure each prediction is highly personalized.
“What we’d like to put together is a model that captures the bigger context of what a person is experiencing […] and try to make predictions that a person is going to care about,” said Pujara.
To illustrate the role of context, Pujara explains the example of how weather data can inform restaurant recommendations. On a warm spring day, the AI system may recommend a restaurant with outdoor patio seating for one user, while avoiding this recommendation for another user who has allergies. Selective decision-making processes like these may seem like a no-brainer for most of us, but for AI models, the effect of context on decision-making can be very hard to grasp.
Pujara, Getoor, and Wang aren’t first-time collaborators — Getoor was Pujara’s PhD advisor during his time at UCSC, and the pair have worked together on various projects since. Pujara met Wang when he was a visiting student at Carnegie Mellon under William Cohen, their current advisor at Google.
One of the team’s major interests is Probabilistic Soft Logic (PSL), a framework for building probabilistic models that leverage relationships. PSL makes AI capable of handling levels of uncertainty and flexibility that remain challenging for machine learning systems. In essence, these recommender systems would be able to predict preferences and make suggestions based on a complex combination of contextual data, relationship data, and reasoning.
“The goal of PSL is to capture logical relationships between things in a way that takes into account probability uncertainty — that human view of the world where things aren’t “if, then” but “if, probably’,” Pujara explained.
In this project, Pujara and his team are researching how to best achieve a more cognizant AI model by using dialogue training. Currently, AI bots such as Siri and Amazon’s Alexa are useful for our informational needs, but they can’t connect with or understand us on a more personal level like other humans can. Using PSL as a framework, dialogue with AI bots is given the potential to be more impactful and meaningful.
Currently, Pujara is working with a cohort of USC students to further flesh out the prospect of an understanding and empathizing AI bot. But in order to achieve the creation of a human-like AI bot, there’s a lot of work required in deciphering the implicit information that we often take for granted in human interactions.