Despite increasing computer power, machines still struggle to learn visual tasks as easily as humans, according to a computer scientist who spoke at a recent lecture at Ateneo de Manila University.
At the second Ateneo Breakthrough Lecture held on February 26 in Escaler Hall, Dr. Patricia “Pi” Angela R. Abu explained how machine learning systems require much more structured training than humans to interpret images and visual environments.
“One of the most surprising things about machine learning is that, despite how powerful computers are, they don’t learn the way humans do. Young children can easily recognize familiar faces, tell when something looks unusual, and make sense of a crowded playground with little instruction, but these same tasks can be difficult and exhausting for computers,” Abu said.
He noted that computer vision systems typically require large datasets, extensive labeling, iterative training, and continuous testing to function reliably under a variety of conditions, such as changes in lighting, camera angles, weather, and other real-world variables.
Mr. Abu gave a talk entitled “Smarter Sight: Building Intelligent Visual Systems for Public Good.” The talk explored the gap between human and machine cognition. This means that machines can sometimes outperform humans at analyzing large numbers of images, but they require far more preparation and training.
Abu, an associate professor and dean of the Department of Information Systems and Computer Science (DISCS) at Ateneo de Manila University, leads the university’s ALIVE research group focused on machine learning, computer vision, and image processing systems.
Her team’s projects include applications in healthcare and transportation. In the medical field, ALIVE has developed tools such as dental image support systems and deep learning models for detecting bone metastases from medical scans.
The group has also developed V-PROBE (Vehicle and Pedestrian Real-Time Observation and Behavior Assessment), a system designed to monitor traffic flow, predict parking availability, and identify potential congestion before it worsens.
Abu said such projects require close collaboration with organizations that manage complex real-world environments, as machine learning models need to run not only in controlled demonstrations but also under the unpredictable conditions of day-to-day operations.
The research team is now exploring deeper collaborations with industry partners to help test and deploy these technologies outside the lab, she said. These partnerships can provide real-world production configurations, data pipelines, and deployment paths needed to evaluate systems for speed, privacy and security protection, hardware limitations, and reliability in a variety of conditions.
This talk is part of Ateneo’s Breakthrough Series, which focuses on research and development and emerging technologies with potential applications for the public good.
