Face detection in untrained deep neural networks

This paper investigates whether face-selectivity in the brain arises innately or requires training. Using a hierarchical deep neural network model of the ventral visual stream, the authors demonstrate that face-selectivity emerges robustly in randomly initialized networks, reproducing characteristics observed in monkeys and enabling face detection. Interestingly, selectivity to other non-face objects also emerges innately. This suggests that random feedforward connections in untrained deep neural networks may suffice to initialize primitive visual selectivity.

Seungdae Baek, Min Song, Jaeson Jang, Gwangsu Kim, Se-Bum Paik