1. Exploring the role of natural statistics in helping us make sense of the world
Our prior knowledge about how the world looks like can help our visual system making sense of noisy or ambiguous information. In this line of research, I seek to understand the role of statistical regularities in aiding object representations, by using a combination of neural and computational approaches.
Forthcoming talk at virtual-VSS, May 2021).
2. What is the nature of representational formats in brains and deep neural networks?
In this line of research, I am interested in relating neural and machine representation to different similarity spaces, to better understand the format that object representations take as they progress through the visual hierarchy.
Here's a video of a talk I gave on this project (virtual-VSS, June 2020).
3. Investigating the organizing principles underling object representations in high-level visual cortex
Inanimate object representations in the occipitotemporal cortex follow a large-scale organization along the dimensions of real-world size (Konkle and Oliva, 2012). But, what do we mean by size? Do neural responses reflect 1) an abstract interpretation of a stimulus as being big or small, or 2) does it reflect something more visual in nature, such as the typical shapes of big and small objects?
To explore the first possibility, we tested how real-world size organization could be alternatively explained by the property of motor-relevance. Indeed, motor-relevance is a high-level, non-perceptual property that covaries with size: small objects tend to be more often hand-held and manipulated, while big objects tend to be less so.
To explore the second possibility, we related real-world size to a mid-level visual feature that correlates with it in the real world: curvature. In the world, small objects tend to be curvier, while big objects tend to be boxier. It is possible that curvature alone might drive the organization by size of inanimate objects in the ventral stream; alternatively, real-world size information might be preserved even when correcting for curvature. We tested this in a 2 x 2 behavioral and fMRI design.
You can see what we found here (VSS 2019).