Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Competitive Retraining Reveals the Resilience and Reallocation of Functional Specialization in Deep Neural Networks

Zhengqing Miao¹, Katharina Dobs¹; ¹Justus Liebig Universität Gießen

Presenter: Zhengqing Miao

Cognitive neuroscientists have long documented functional specialization in the brain for tasks such as face, body or scene recognition, and recent computational studies reveal that deep neural networks (DNNs) spontaneously develop specialized populations of units for the same tasks. But are these specialized units necessary for performance, and how plastic are they? Here, we combine lesioning approaches with competitive retraining in DNNs to address these questions. In a dual-task network with localized specialized units in the final convolutional layer for face and object tasks, we ablated those units either at the onset or continuously throughout retraining. To modulate competition, we retrained the networks on a single task or both tasks simultaneously. Our findings reveal that retraining restores network performance even when these layer-specific units remain permanently disrupted, indicating they are not strictly necessary. Moreover, the extent and pattern of unit reallocation vary with retraining conditions, demonstrating substantial plasticity and suggesting that the reallocation process is an intrinsic outcome of rapid network optimization.

Topic Area: Object Recognition & Visual Attention

Extended Abstract: Full Text PDF