Research
I study how cognitive representations emerge over development in biological and artificial intelligence. My research asks how early constraints and experience shape later learning, how structured knowledge develops over the course of learning, and how we can interpret and control cognitive structure within modern AI systems. Across these directions, I am especially interested in how seemingly limited early conditions can provide useful inductive biases that support robust, abstract, and human-aligned representations.
I investigate how neural networks can be better prepared for learning before they encounter real sensory input from the external world. Inspired by early biological development, I study both inductive biases that emerge prior to training and learning processes driven by spontaneous, unstructured activity. This work treats initialization not simply as a numerical starting point, but as an early developmental phase that can shape the efficiency, robustness, and reliability of subsequent learning.
Selected Publications
J. Cheon, S.-B. Paik.
Brain-inspired warm-up training with random noise for uncertainty calibration
Nature Machine Intelligence 8, 602–613 (2026)
* Featured in News & Views: T. Isomura, Learning to be uncertain before learning from data
J. Cheon, S. W. Lee, S.-B. Paik.
Pretraining with random noise for fast and robust learning without weight transport
Advances in Neural Information Processing Systems 37, 13748–13768 (NeurIPS 2024)
J. Cheon*, S. Baek*, S.-B. Paik (* Co-first).
Invariance of object detection in untrained deep neural networks
Frontiers in Computational Neuroscience 16, 1030707 (2022)
Unlike artificial vision models, which are typically trained on high-fidelity images from the outset, humans begin learning with limited sensory capacities that gradually mature over development. I study how these early visual constraints shape later perceptual and representational development. I incorporate developmentally inspired sensory limitations into neural networks to test how they can promote more structured and abstract representations, and use neural network models to understand atypical visual development in late-sighted children.
Selected Publications
J. Cheon, M. Vogelsang, L. Vogelsang, P. Sinha.
Early visual experience scaffolds hierarchical visual representations
Advances in Neural Information Processing Systems 39 (NeurIPS 2026)
J. Cheon, S.-B. Paik.
Gradual sensory maturation promotes abstract representation learning
bioRxiv Preprint 2025.06.24.661295 (2025)
L. Vogelsang, M. Vogelsang, J. Cheon, P. Gupta, P. Shah, P. Sethi, S. Narang, S. Ganesh, P. Sinha.
Compromised Gestalt grouping in late-sighted children
Manuscript in preparation
I investigate how core knowledge and structured representations can emerge when neural networks learn from the naturalistic visual experience available to young children. Using child-centered egocentric video, I study how category structure and core visual competencies develop over time, and how early perceptual constraints can serve as useful inductive biases for grounded word acquisition. More broadly, this work asks how structured knowledge can emerge from developmentally realistic experience.
Selected Publications
J. Cheon, M. Vogelsang, L. Vogelsang, P. Sinha.
Early visual degradation facilitates few-shot word learning from a child's egocentric input
Developmental Perspectives on AI Workshop @ NeurIPS 2026
C. Shin, J. Cheon, B. M. Lake.
Coarse-to-fine category structure in child's-view model representations
Developmental Perspectives on AI Workshop @ NeurIPS 2026
I study how subspace aligned with human cognition is organized within modern foundation models and how these internal representations can be causally manipulated to shape model behavior. My work examines how semantic abstraction is geometrically embedded in vision-language and large language models, and whether intervening on these structures can enhance abstraction-related behavior. I also develop interpretability methods to uncover cognitively meaningful organization within model representations. More broadly, I aim to use these approaches both as computational models for studying human cognition and to develop safer and more controllable AI.
Selected Publications
J. Cheon, M. Vogelsang, L. Vogelsang, P. Sinha.
Spectral stratification of semantic abstraction in vision-language models
Advances in Neural Information Processing Systems 39 (NeurIPS 2026)
* Also presented in Mechanistic Interpretability Workshop @ ICML 2026
J. Cheon, T. Fel.
Discovering hierarchical concept geometry with sparse autoencoders
Manuscript in preparation