research

I study how the visual input children actually receive shapes the object representations they build, and how that input compares to what vision models are trained on—using large-scale egocentric video, computer vision, and behavioral experiments.

Characterizing infant visual experience

What do young children actually see, and how does it compare to the data that drives modern vision systems? I use naturalistic egocentric video (the BabyView dataset) to quantify objects in the child’s view and compare those statistics to curated datasets such as THINGS.

• Variability in young children’s everyday visual experiences of object categories (CCN 2026, spotlight) — Global and local dispersion over 85 categories and 7,018 validated crops, separating viewpoint-driven from instance-driven variability. Poster · Interactive explorer.
• Characterizing the visual representation of objects from the child’s view (under review) — Object frequency and representational geometry across 868+ hours of egocentric video from 31 families, compared with THINGS using CLIP and DINOv3.
• Characterizing the inputs to infants’ object category representations (VSS 2026) — How these inputs relate to early object categories.
• Quantifying infants’ everyday experiences with objects in a large corpus of egocentric videos (CCN 2025) — Automated object detection over 868+ hours of infant headcam video.

Visual–linguistic alignment

How well are the visual and linguistic streams aligned in naturalistic settings? I use multimodal models (e.g., CLIP) to measure that alignment in egocentric infant data.

• Assessing the alignment between infants’ visual and linguistic experience using multimodal language models (Tan, Yang, et al., 2025) — Vision–language models to assess alignment over time and what makes naturalistic multimodal input learnable.

Attention, action, and learning

How do attention and action structure learning—e.g., how manual actions create visual saliency and support joint attention, and how real-time attention relates to language?

• Toddlers’ active gaze behavior supports self-supervised object learning (Yu, Aubret, Raabe, Yang, Yu, & Triesch, Developmental Science, 2026) — Toddlers’ gaze during dyadic play supports self-supervised learning of view-invariant object representations.
• Using manual actions to create visual saliency (CogSci 2023) — Manual actions create visual saliency and support joint attention from an “outside-in” perspective.
• Learning semantic knowledge based on infant real-time attention and parent in-situ speech (CogSci 2024) — Linking infant gaze and parent speech to ask how attention shapes which semantic information is available during learning.

3D shape and object experience

Do children and vision models converge on 3D object shape? I built MOCHI Kids, a developmental (ages 4–6) adaptation of the MOCHI shape-oddity benchmark, for direct human–model comparison on multiview object consistency. The web experiment (TypeScript, jsPsych, Express, MongoDB) is in piloting. In parallel, I am building a synchronized capture rig—head-mounted eye tracking, depth cameras, and a room camera—to measure how action and dyadic interaction change the object views children get.

Methods and open science

Methods: Computer vision (object detection, pose estimation, multimodal embeddings), multimodal data fusion (head-mounted eye trackers, cameras, microphones), and behavioral experiments (web and in-lab, including human–model benchmarks).
Open science: Reproducible pipelines for large-scale naturalistic data, including a BabyView Objects reproducibility capsule, a CCN 2026 object-variability analysis package, and a layer-wise embedding explorer.