Artificial general intelligence elements for enhancing biometric computer vision algorithms

Abstract

The contemporary advancements in Artificial Intelligence (AI) have primarily focused on narrow, specialised applications. Yet, these models often fail catastrophically under unexpected conditions and require significantly more computational resources than human intelligence. This research addresses significant gaps in existing machine learning (ML) methodologies by integrating principles of Artificial General Intelligence (AGI), specifically emulating human cognitive faculties such as self-awareness, adaptability, proactive agency, memory-based learning, and imagination. The research focuses explicitly on biometric applications within Computer Vision, particularly Face Recognition (FR) and Facial Expression Recognition (FER). An original Supervisor Artificial Neural Network (SNN) was developed to introduce self-awareness by assessing the uncertainty and trustworthiness of existing stateof-the-art (SOTA) convolutional neural network (CNN) predictions. Experimental results demonstrated substantial improvements, increasing accuracy by the SNNaugmented CNN models compared to purely CNN models by up to 15% for FR tasks and 30–40% for FER tasks, particularly under challenging conditions involving makeup and occlusions previously unseen by the models. Further, the introduction of a statistical loss function with memory allowed dynamic adjustment of trustworthiness thresholds based on historical successes and failures, increasing accuracy by an additional 10%. Active learning methods were also applied, proactively requesting targeted additional training data to mitigatemodel uncertainty, further enhancing performance by up to 10% in FR tasks. Lastly, to address data scarcity in FER tasks, a generative ANN employing an encoder-decoder architecture with advanced cosine distance attention mechanisms and plasticity learning functions was proposed. The generated synthetic data augmented the training set, resulting in a 5–10% performance boost on complex FER datasets. This approach not only improves the accuracy, robustness, and resource efficiency of biometric ML models but also generalises effectively, indicating significant potential for broader application across various AI domains.

Date of Award	13 May 2024
Original language	English
Awarding Institution	University of Bedfordshire
Supervisor	Vitaly Schetinin (Supervisor) & Tess Crosbie (Second supervisor)