Researchers at Penn State have unveiled a groundbreaking approach to train artificial intelligence systems to recognize objects and navigate their surroundings, paving the way for the development of more advanced AI technologies that can explore extreme environments or distant planets.
During the initial two years of life, children rapidly develop expertise in recognizing a limited range of objects and faces, but only under a multitude of varying lighting conditions and viewpoints. Researchers were impressed by this groundbreaking developmental insight, which led them to pioneer a novel machine learning approach leveraging spatial context information to significantly enhance AI visual model training efficiency. Researchers found that AI models trained using the novel approach significantly outperformed their baseline counterparts, achieving a notable improvement of up to 14.99%. The researchers published their results in the inaugural issue of the journal Pattern.
“Current techniques in AI rely on massive datasets comprising randomly curated and shuffled images sourced from the internet for training purposes.”
Our technique is informed by developmental psychology, which examines how children understand the world, said Lizhen Zhu, the lead researcher and doctoral candidate in the School of Information Sciences and Technology at Penn State.
Researchers devised a novel contrastive learning algorithm, a self-supervised approach whereby AI systems learn to identify visual patterns and differentiate between images that share a common origin, yielding a positive pair in the process. While these algorithms often treat images of the same object captured from various viewpoints as individual entities rather than as constructively paired pairs? By accounting for environmental context and location, the AI system can effectively overcome these obstacles and identify relevant pairings despite changes in camera position or orientation, lighting conditions, angle or situation, and focal length or zoom, as revealed by the researchers.
“We propose that infants’ observable gaze allocation is influenced by their understanding of spatial layout.” To create a self-contained dataset featuring spatiotemporal patterns, we leveraged the capabilities of the ThreeDWorld platform, configuring virtual environments that simulate real-world scenarios with unprecedented realism and interactivity. With this capability, we could precisely calibrate and monitor the positioning of surveillance cameras as if a child were innocently walking through a residence, noted Zhu.
Researchers designed a trio of simulated environments: House14K, House100K, and Apartment14K, which respectively contained around 14,000 and 100,000 images of various patterns, accurately recreating realistic settings for testing purposes? They subsequently executed baseline contrasting methods and novel algorithms through simulations three times, examining how each classified images effectively. Staff found that models trained on their algorithm surpassed lower-performing models on numerous tasks. In terms of recognizing the room within the digital condominium, the enhanced model achieved a remarkable 99.35% accuracy rate, representing a significant 14.99% improvement over its baseline counterpart. The datasets are now available for various scientists to utilize in mentoring.
Learning new fashion trends in an unfamiliar environment with limited knowledge can be extremely challenging. According to James Wang, a distinguished professor of information sciences and technology and advisor to Zhu, the team’s work marks one of several pioneering attempts to harness AI training through visual content for enhanced energy efficiency and versatility.
The findings have significant implications for the development of advanced AI techniques capable of learning from and adapting to novel environments, as requested by the scientific community.
“When navigating uncharted territories with limited resources, autonomous robots may find this approach particularly effective,” Wang said. “To set the stage for future developments, we intend to enhance our model by leveraging spatial data more effectively and incorporating a broader range of environments.”
Collaborators from Penn State’s Division of Psychology and the Division of Computer Science and Engineering also contributed to this study. The research reported in this manuscript was supported by a grant from the United States Government. The Nationwide Science Foundation collaborates with the Institute for Computational and Informatics Sciences at Pennsylvania State University.
