Nicholas Turk-Browne

Nicholas Turk-Browne

Yale University
1 previous affiliation
1Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Nicholas Turk-Browne focuses on the functional architecture of the human visual system, the development of cortical organization in infancy, and the computational modeling of visual learning processes. His work integrates functional magnetic resonance imaging (fMRI) with machine learning and computational cognitive science to understand how the brain represents visual information and how these representations evolve through experience and maturation.

Research Trajectory

The available evidence from 2022 and 2023 indicates a research program that bridges high-level computational modeling with empirical developmental neuroscience. In 2022, Turk-Browne participated in collaborative efforts to refine the relationship between artificial intelligence and human vision. This included work on establishing benchmarks to compare unsupervised learning algorithms with human visual learning capabilities across different timescales. During this period, he also contributed to the development of new methodological frameworks for neuroimaging, specifically focusing on how to extract shared neural structures from noisy, high-dimensional fMRI data across multiple subjects.

By 2023, Turk-Browne’s work as a senior investigator shifted toward the fundamental question of how the visual cortex is organized at the earliest stages of human life. This research utilizes advanced fMRI techniques to map the infant brain, providing evidence for the early functional maturity of the visual system. This trajectory suggests a move from broader participation in computational vision and neuroimaging methodology toward a leadership role in developmental cognitive neuroscience, specifically investigating the presence and precision of cortical maps in the infant brain.

Current Directions

A primary focus of Turk-Browne’s current research involves the use of fMRI to investigate the retinotopic organization of the visual cortex in human infants. This work addresses a long-standing debate in neuroscience regarding the extent to which functional maps in the striate and extrastriate cortex are present at birth or emerge through prolonged visual experience. By applying retinotopic mapping techniques to infants aged 5 to 23 months, his research seeks to characterize the boundaries of visual areas V1 through V4 and the gradients of spatial frequency preferences that define these regions.

Additionally, Turk-Browne continues to engage with the computational constraints of visual learning. This involves evaluating how well current machine learning models—particularly those utilizing unsupervised learning—can replicate the rapid, flexible learning seen in humans over short intervals (real-time) and the robust accumulation of knowledge over longer periods (life-long). This line of inquiry aims to identify the gaps between biological and artificial vision, potentially informing the development of more human-like computer vision systems.

Representative Work

In Retinotopic organization of visual cortex in human infants, Turk-Browne served as the senior author on a study investigating the functional maturity of the infant visual system. The research utilized fMRI to perform retinotopic mapping in infants between 5 and 23 months of age. The study identified clear retinotopic maps with alternating preferences for vertical and horizontal meridians, which serve as indicators for the boundaries of visual areas V1, V2, V3, and V4. Furthermore, the work revealed an orthogonal gradient of preferences for spatial frequencies, ranging from high to low. These findings suggest that the functional maps organizing the adult visual cortex are already present and operational in early infancy, providing a structural foundation for rapid visual development.

Turk-Browne also participated in research focused on the intersection of machine learning and human cognition in How Well Do Unsupervised Learning Algorithms Model Human Real-time and Life-long Learning?. This work addresses the challenge of modeling the human ability to learn from visual streams at multiple timescales. The study established benchmarks for real-time and life-long continual visual learning to measure how closely unsupervised algorithms can match human behavior. By evaluating models on their ability to adapt to rapid changes in visual input over minutes and hours, the research highlights the current limitations of computational models in replicating the flexibility and robustness of human visual knowledge acquisition.

In the domain of neuroimaging methodology, Turk-Browne contributed to the development of new techniques for analyzing multi-subject data in Learning Shared Neural Manifolds from Multi-Subject FMRI Data. This work addresses the problem of high dimensionality and noise in fMRI signals. The researchers proposed a manifold-regularized multiple-decoder autoencoder (MRMD-AE) designed to learn a common embedding space across different brains. By assuming that stimulus-driven responses share latent features common across subjects, this neural network approach aims to denoise fMRI signals and capture the intrinsic structure of neural data more effectively than previous linear methods like principal component analysis or standard shared response modeling.

Profile updated

See something wrong? Report a mistake