Idris Nemsia | 2025 I.S. Symposium

Name: Idris Nemsia
Title: One-Shot Facial Recognition using Siamese Neural Networks
Major: Computer Science
Advisor: Sofia Visa

Traditional deep learning models, particularly Convolutional Neural Networks (CNNs), have demonstrated significant success in facial recognition tasks. However, these models encounter limitations when faced with insufficient data or when classifying rare classes. Few-shot learning has emerged as an effective solution to address these challenges. It is especially valuable in scenarios with limited data or when the environment is dynamically changing, as it is less resource-intensive and does not require retraining each time the data changes. This thesis explores the application of Siamese Neural Networks (SNNs) for facial classification, leveraging one-shot learning. An SNN model was trained using pairs from the Olivetti Faces dataset, achieving an accuracy of 68.07% on 10-way one-shot trials, where it classified unseen classes based on a single image per class. The model’s generalization capabilities were further tested on the Yale Face Database without retraining or fine-tuning, where it achieved a classification accuracy of 44.27% in a similar 10-way one-shot scenario. These results highlight the SNN’s ability to adapt to extreme scenarios with minimal data and still produce accurate results. However, they also emphasize the need for more diverse and complex datasets to improve generalization and reliability. Future work will focus on enhancing the model by exploring more diverse datasets and optimization techniques, such as modifications to loss functions and the application of optimization strategies like data augmentation.

Posted in Symposium 2025 on May 1, 2025.