Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Separating Numerosity and Proportion Representations Through Pretraining

Yechan Cho¹, Hyeonsu Lee¹, Emma Deneuville², Jaeson Jang¹; ¹Korea Institute of Science and Technology, ²Université Clermont Auvergne

Presenter: Yechan Cho

Numerosity and proportion are two fundamental types of quantitative information extracted from visual scenes, and both can be discriminated early in life. Although these cues are correlated in natural environments, incongruent combinations can occur. Young children can distinguish each individually but struggle when they conflict, suggesting that the neural representations supporting them are not yet fully refined. Here, we tested whether training on natural image classification alone can promote this refinement. We compared an untrained neural network with a network pre-trained on natural images and analyzed their latent representations. Pre-training increased the principal angle between subspaces and reduced interference in ablation analyses. Consistently, the pre-trained network showed higher accuracy in an interference task. These findings suggest that natural visual experience reorganizes quantitative representations into a more separable.

Topic Area: Computational Models of Vision & Visual Cortex