Robust Visual Perception
A key characteristic of human visual perception is its robustness. Despite variations in the external environment, our internal visual representations remain stable and consistent. What mechanisms enable this robustness, and how can they be understood? Answering this question is also crucial for developing reliable machine vision systems.
Imagine a scenario where a person is driving under adverse weather conditions. The driver must navigate through traffic, avoid other vehicles, and be wary of pedestrians, all while contending with significant visual disruptions like heavy rain or snow. Alternatively, consider the situation of finding a friend in a crowded place. Despite the presence of numerous individuals, identifying a familiar face is not as challenging as it might seem. Humans possess an exceptional ability for robust object recognition, a core component of our visual perception, which facilitates seamless interaction with our surroundings despite obstacles such as variations in size and orientation, differences in lighting, or even partial occlusions.
Over decades, the field of vision science has endeavored to uncover the mechanisms behind the robust nature of human vision. This research has grown in importance, especially since deep learning models, which are often said to perform as well as or better than humans in various visual tasks, frequently struggle in practical applications. Examples of these shortcomings include self-driving vehicles failing to adapt to novel scenarios, surveillance systems inaccurately identifying faces, and medical imaging technologies making erroneous assessments.
Our research group is dedicated to leveraging diverse psychological methodologies, neuroimaging technologies, and deep learning models to decipher the secrets behind the robust nature of human vision. Notably, deep learning models offer a groundbreaking method to validate numerous hypotheses within neuroscience. Specifically, the comparative analysis between human and machine vision provides insights into how visual information is internally represented in both, presenting new perspectives on the essence of robustness in visual perception. We believe our research efforts will have profound implications for both neuroscience and the development of safer, more advanced AI technologies.
Image from Jang and Tong, Nature Communications, 2024
