Improving Anomaly Detection in the Cognitive Test Scores using Siamese Neural Network and Metric Learning as Ordinal Classification Task

Abstract

The detection of outliers in similarity measurements is a crucial challenge in machine learning, particularly for image-based applications such as medical diagnostics. Existing methods for outlier identification using Siamese networks often lack comprehensive experimentation on the impact of various hyperparameters. This study addresses this gap by: • Conducting extensive experiments to analyze the behavior of the Siamese network under different hyperparameter configurations. • Utilizing pretrained ResNet101 and ResNet50 architectures as the backbone CNNs, combined with contrastive loss and a customized dataset generation approach. • Demonstrating significant improvements, with the two proposed models achieving a false positive rate of 1.96% and reducing false positives by 1.94% compared to existing systems. • Achieving these results using only 2000 image pairs and within just 50 epochs, showcasing the robustness of the approach with minimal training data. This work provides a detailed analysis of hyperparameter optimization in the convolutional neural network of the Siamese network architecture for anomaly detection in cognitive test scores. The findings have the potential to enhance the performance of convolutional neural networks within the Siamese framework, offering a valuable contribution to advancing outlier detection methodologies.