|
A Neuro-Symbolic Framework for Face Recognition with a SOAR-Based Cognitive Meta-Controller, Parameter-Efficient Transfer Learning (PEFT), and Test-Time Adaptation (TTA)
|
Zinat Mehdidoust Jalali   , Azamossadat Nourbakhsh * , Sara Motamed |
| Department of Computer Engineering, La.C., Islamic Azad University, Lahijan, Iran , Nourbakhsh@iau.ac.ir |
|
|
Abstract: (34 Views) |
Face recognition in real-world environments remains challenging due to domain shifts, limited labeled data, decision uncertainty, and constrained computational resources. This paper presents a novel neuro-symbolic framework that combines a SOAR-based cognitive meta-controller, cost-aware reinforcement learning, and domain adaptation to jointly improve recognition accuracy, computational efficiency, and inference latency. A face embedding network is first trained on the source domain and then adapted to the target domain using parameter-efficient fine-tuning (PEFT) and Domain-Adversarial Neural Networks (DANN). To enhance robustness under limited supervision and distribution shifts, the framework further incorporates few-shot learning and test-time adaptation (TTA). The proposed meta-controller is formulated as a Markov Decision Process (MDP), enabling dynamic allocation of computational resources based on image quality, uncertainty estimation, and recognition status. Symbolic knowledge encoded in the SOAR architecture guides the reinforcement learning policy through a neuro-symbolic gating mechanism, improving both interpretability and decision consistency. Experiments on the IJB-C and MegaFace benchmarks demonstrate significant improvements in face recognition and open-set recognition performance, while PEFT reduces trainable parameters by more than 95% without increasing inference latency. Overall, the proposed framework offers an accurate, adaptive, interpretable, and computationally efficient solution for real-world face recognition applications. |
|
| Keywords: Face Recognition, Neuro-Symbolic Artificial Intelligence, SOAR Cognitive Architecture, Parameter-Efficient Fine-Tuning (PEFT), Test-Time Adaptation (TTA), Reinforcement Learning. |
|
|
Full-Text [PDF 937 kb]
(11 Downloads)
|
Type of Study: Original |
Subject:
Other
Received: 2026/05/15 | Accepted: 2026/09/10 | Published: 2026/09/15
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
