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

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Functionally specialized units reflect readout pathways rather than distinct feature encoding in artificial neural networks

Zhengqing Miao¹, Katharina Dobs¹; ¹Justus Liebig Universität Gießen

Presenter: Zhengqing Miao

Functional specialization, in which neurons or units causally contribute to specific tasks, emerges in both biological visual systems and artificial neural networks (ANNs). In ANNs, it is often assumed that task-critical units encode dedicated features. Yet causal involvement does not necessarily imply unique feature encoding: such units may instead function as preferred readout pathways over a shared feature space. To dissociate between these accounts, we use intrinsic spatial frequency (SF) tuning as an interpretable proxy for representational structure and perform systematic lesion analyses in a dual-task VGG network trained on face and object categorization. We find that identified face- and object-critical units show highly similar SF tuning, suggesting they do not rely on strongly distinct feature representations. While permanently lesioning these units severely impairs task performance, the network's global SF profile remains largely unchanged. Critically, performance can be fully restored by fine-tuning only the fully connected layers, indicating that such units are not necessary for performance. These findings support the view that functional specialization in ANNs reflects selective downstream readout over a shared feature space, rather than dedicated encoding of category-specific features.

Topic Area: Computational Models of Vision & Visual Cortex