Efficient Deep Discriminant Embedding: Application to Face Beauty Prediction and Classification

Efficient Deep Discriminant Embedding: Application to Face Beauty Prediction and Classification Fadi Dornaika1,2, Abdelmalik Moujahid1, Kunwei Wang3, Xudong Feng3 1University of the Basque Country UPV/EHU, San Sebasti´an, Spain 2IKERBASQUE, Basque Foundation for Science, San Sebasti´an, Spain 3Northwestern Polytechnical University (NPU), Xian, China Published in: Engineering Applications of Artificial Intelligence, Volume 95, 2020, Article 103831, Elsevier. DOI: 10.1016/j.engappai.2020.103831. Available online 1 August 2020. Abstract Inspired by deep learning architectures, we introduce a multi-layer local discriminant embedding algorithm that integrates feature selection as a main step to capture the most relevant and discriminant features of an input face image or face descriptor. The proposed framework allows to transform any linear method to a deep variant via a cascaded feature extraction and selection architecture able to convert weak and noisy descriptors to strong ones. As a case study, the local discriminant embedding (LDE) projection is adopted as a linear feature extraction method. The resulting framework can be considered as an efficient deep discriminant embedding technique. To validate this framework, we have considered two different computer vision problems: face beauty prediction which involves both classification and regression tasks, and face recognition which is a classical classification problem. Experiments conducted on different public benchmark databases show that this approach enhances the performance of the LDE algorithm and provides a discriminating strategy to solve the dimensionality reduction problem. For face beauty regression, our proposed framework achieved on average an improvement of about 5% and 7% with respect to two other configurations where only VGG-face and VGG-face followed by LDE have been considered. For face beauty classification, the proposed algorithm outperformed many classical manifold learning techniques reaching in some databases improvements of about 10%. Main Contributions •Proposes a multi-layer cascaded architecture combining local discriminant embedding and Fisher-based feature selection, yielding efficient deep metric learning embeddings. •Extends manifold learning by integrating sequential feature cleaning steps, boosting class separation and compactness in face beauty and recognition problems. •Achieves notable improvements (5–10% accuracy) in classification and regression tasks over state-of-the-art methods on multiple face beauty and recognition datasets. •Demonstrates the method’s robustness with a low number of parameters and training samples, facilitating learning without large data or expensive GPU resources. 1 •Provides flexibility to substitute manifold learning modules, supporting alternative algorithms beyond LDE, and validates impact across different descriptors and validation settings. Impact of the Paper The proposed efficient deep discriminant embedding algorithm marks a significant advance in the intersection of manifold learning, feature selection, and deep architecture design for artificial intelligence. Its layered integration of linear embedding, discriminative metric learning, and recursive feature cleaning provides not only a methodological improvement over shallow approaches, but also a practical framework for scenarios with limited data and computational resources. Key aspects of the impact include: •Bridging deep learning and manifold learning: Traditionally, deep representation learning relies primarily on neural networks. This work demonstrates that classical manifold learning algorithms (e.g., LDE) can be recursively integrated in a “deep” fashion, bringing many advantages of deep architectures—such as abstraction, denoising, and robust feature hierarchies—to tasks where neural networks may not be feasible or optimal. •Versatile application domain: The methodology is validated across both classification and regression settings, notably in facial beauty prediction and face recognition, but is applicable to any problem where feature compactness, interpretability, and discriminative power are essential, including medical imaging, biometrics, industrial inspection, and more. •Enabling learning with scarce labels: By relying on iterative feature selection and embedding, the proposed method achieves excellent performance with fewer parameters and less labeled data than typical end-to-end deep neural networks—a critical factor for applications with limited annotated samples or where interpretability is mandated. •Enhancing explainability and efficiency: The results reveal that with appropriate feature selection and metric learning, “deep” discriminant embeddings can offer high performance without the black-box limitations of standard neural networks. This enhances trust and transparency for AI adoption in regulated or mission-critical environments. •Catalyzing scalable AI: This work encourages a shift toward hybrid architectures that reuse and extend classic algorithms within scalable, multi-layer structures. It makes deep learning concepts more accessible for a broader range of researchers and practitioners by reducing the computational and data costs typically required for state-of-the-art results. •Foundation for future research: The proposed framework can be adapted and extended—by replacing the linear module, adjusting selection strategies, or stacking with other discriminant embedding methods—paving the way for further innovations in metric learning, transfer learning, and interpretable AI models. Overall, this paper demonstrates that efficient deep discriminant learning is possible without large-scale neural networks, and that systematic integration of feature selection and manifold embedding sets a new standard for compact, high-performing, and interpretable AI in many vision and data-rich domains. Reference: Fadi Dornaika, Abdelmalik Moujahid, Kunwei Wang, Xudong Feng. “Efficient deep discriminant 2 embedding: Application to face beauty prediction and classification.” Engineering Applications of Artificial Intelligence, Volume 95, 2020, Article 103831. DOI: 10.1016/j.engappai.2020.103831 3