Projects
Ensemble Representations in Physical Reasoning
We investigate how people perceive physical properties of groups (ensembles) in complex scenes, extending the concept of ensemble perception to intuitive physics. Through experiments comparing individual and group mass judgments, we find that people excel at processing ensembles beyond the sum of individual properties, suggesting abstract representations of group-level physical attributes.
Examples of lighter and heavier stimuli and the task setup for the ensemble perception experiment. Participants compared the average mass of 25 marbles falling onto a cloth to a similar group with five changed marbles, then identified one changed marble. Their mass judgments exceeded what individual processing would predict based on their localization accuracy, supporting ensemble perception in intuitive physics.
Probabilistic Reasoning in Children
Recent evidence suggests that while young children can evaluate probabilities in observable populations, they often struggle with reasoning about possibilities. To investigate this, we designed experiments to explore how children and adults simulate potential outcomes in probabilistic reasoning tasks, revealing a developmental divide in their mental simulation strategies.
To test how children reason about possibilities, we designed a task that involved catching a ball that could fall into one of six bins, with rooms of increasing uncertainty. Participants, children and adults, positioned three "cushions" to catch the ball, either spreading them across bins (indicating awareness of multiple possibilities) or stacking them in one bin (suggesting limited understanding of possibilities).
GeoCLIP: Worldwide Image Geo-localization
GeoCLIP is a CLIP-inspired model that aligns images and geographic coordinates in a shared embedding space. By learning a continuous representation over the Earth, it treats each location as having its own place embedding rather than assigning images to a fixed set of geographic categories. This establishes a more flexible framework for worldwide geo-localization and re-imagines geography as a structured representational space.
GeoCLIP learns a shared representation for images and coordinates, turning the Earth into a continuous latent field of place embeddings.