Vision models & the ventral stream

This project investigates the relationship between visual learning objectives, representation geometry, and biological vision. The resulting work, “Vision Models Trained to Estimate Spatial Latents Learn Ventral-Stream-Aligned Representations,” appeared at ICLR 2025.

My contributions focused on model analysis, representation similarity experiments, and hypothesis testing. I implemented CKA and RSA methods and used them to test relationships between architectural inductive biases and learned representations.

My contribution

• Implemented centered kernel alignment (CKA) and representational similarity analysis (RSA).
• Tested hypotheses linking representation geometry to architectural inductive biases.
• Contributed model analysis and representation similarity experiments to the ICLR 2025 paper.