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

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Natural image training enhances generalization in visual number sense of deep neural networks

Joohyun Lee¹, Joonkoo Park², Hansem Sohn¹; ¹Sungkyunkwan University, ²University of Massachusetts at Amherst

Presenter: Joohyun Lee

Number sense, the intuitive ability to estimate the number of items in a scene, has been observed in diverse animal species. Recent studies have shown that this ability also emerges in deep neural networks, yet the inductive biases that drive its emergence remain poorly understood. Here, we investigate how different training curricula shapes the number sense and its robustness in deep learning models. We tested two curricula: one set of models trained de novo on a numerosity estimation task, and the other set pre-trained on natural image categorization. Pre-trained models showed superior generalization to untrained range of numerosity, size, spacing, compared to de novo models. Analysis of models' hidden unit activations revealed that the pre-trained models develop richer and higher-dimensional representational geometry across layers. Together, these results suggest that the statistical properties of natural images provide a powerful inductive bias to develop number sense.

Topic Area: Computational Models of Vision & Visual Cortex