2024년 11월 25일 (월) 오후 4시

연사: 고려대 뇌공학과 장호진 교수님

제목: Leveraging deep neural networks to model human visual representational robustness in challenging conditions

초록: Deep Neural Networks (DNNs) provide valuable insights into the human visual system by allowing comparisons between neural and artificial representational spaces. This talk presents two studies examining how DNNs and human vision respond under suboptimal and noisy conditions. The first study shows that noise-trained networks closely align with human visual processing, predicting recognition thresholds and mimicking human representational robustness in noisy environments. The second study reveals that networks trained on blurred inputs better reflect biological vision, with representations that emphasize shape sensitivity and reduce reliance on high-spatial frequency features. These findings highlight the potential of DNNs to model human visual representations while also highlighting the unique characteristics of biological perception that extend beyond current artificial models.