TY - GEN
T1 - Exploring Trade-offs Between Black-Box and Glass-Box Models in Face Similarity
T2 - 2nd Conference on Human Centered Artificial Intelligence - Education and Practice, HCAI-ep 2024
AU - Jyothi Jayachandran, Abhijith
AU - James, Tonu
AU - Jaiswal, Rajesh
AU - Quille, Keith
N1 - Publisher Copyright:
© 2024 Copyright held by the owner/author(s).
PY - 2024/12/2
Y1 - 2024/12/2
N2 - This study compares black-box and glass-box models for face verification, specifically using Siamese neural networks and K-Nearest Neighbors (KNN). Using the Olivetti dataset, we showed that a Siamese network with VGG19 transfer learning reached 95.4% accuracy, outperforming a custom model's initial 49.61% accuracy. Contrastive loss improved training, while KNN known for its simplicity and interpretability achieved 91.25% accuracy but faced challenges with high-dimensional data. Grad-CAM and saliency maps provided interpretability for the Siamese network and KNN, respectively. This work underscores the trade-off between performance and explainability in these models.
AB - This study compares black-box and glass-box models for face verification, specifically using Siamese neural networks and K-Nearest Neighbors (KNN). Using the Olivetti dataset, we showed that a Siamese network with VGG19 transfer learning reached 95.4% accuracy, outperforming a custom model's initial 49.61% accuracy. Contrastive loss improved training, while KNN known for its simplicity and interpretability achieved 91.25% accuracy but faced challenges with high-dimensional data. Grad-CAM and saliency maps provided interpretability for the Siamese network and KNN, respectively. This work underscores the trade-off between performance and explainability in these models.
UR - https://www.scopus.com/pages/publications/85216573797
U2 - 10.1145/3701268.3701284
DO - 10.1145/3701268.3701284
M3 - Conference contribution
AN - SCOPUS:85216573797
T3 - ACM International Conference Proceeding Series
SP - 62
BT - HCAI-ep 2024 - Proceedings of the 2024 Conference on Human Centered Artificial Intelligence - Education and Practice
PB - Association for Computing Machinery (ACM)
Y2 - 1 December 2024 through 2 December 2024
ER -