Maria Eduarda Caldeira

Abstract
The development of deep learning algorithms has extensively empowered humanity's task automatization capacity. However, the huge improvement in the performance of these models is highly correlated with their increasing level of complexity, limiting their usefulness in human-oriented applications, which are usually deployed in resource-constrained devices. This led to the development of compression techniques that drastically reduce the computational and memory costs of deep learning models without significant performance degradation. These compressed models are especially essential when implementing multi-model fusion solutions where multiple models are required to operate simultaneously. This paper aims to systematize the current literature on this topic by presenting a comprehensive survey of model compression techniques in biometrics applications, namely quantization, knowledge distillation and pruning. We conduct a critical analysis of the comparative value of these techniques, focusing on their advantages and disadvantages and presenting suggestions for future work directions that can potentially improve the current methods. Additionally, we discuss and analyze the link between model bias and model compression, highlighting the need to direct compression research toward model fairness in future works.

Abstract
Clinical diagnosis of syndromes benefits strongly from objective facial phenotyping. This study introduces a novel approach to enhance clinical diagnosis through the development and exploration of a low-dimensional metric space referred to as the clinical face phenotypic space (CFPS). As a facial matching tool for clinical genetics, such CFPS can enhance clinical diagnosis. It helps to interpret facial dysmorphisms of a subject by placing them within the space of known dysmorphisms. In this paper, a triplet loss-based autoencoder developed by geometric deep learning (GDL) is trained using multi-task learning, which combines supervised and unsupervised learning approaches. Experiments are designed to illustrate the following properties of CFPSs that can aid clinicians in narrowing down their search space: a CFPS can 1) classify syndromes accurately, 2) generalize to novel syndromes, and 3) preserve the relatedness of genetic diseases, meaning that clusters of phenotypically similar disorders reflect functional relationships between genes. The proposed model consists of three main components: an encoder based on GDL optimizing distances between groups of individuals in the CFPS, a decoder enhancing classification by reconstructing faces, and a singular value decomposition layer maintaining orthogonality and optimal variance distribution across dimensions. This allows for the selection of an optimal number of CFPS dimensions as well as improving the classification capacity of the CFPS, which outperforms the linear metric learning baseline in both syndrome classification and generalization to novel syndromes. We further proved the usefulness of each component of the proposed framework, highlighting their individual impact. From a clinical perspective, the unique combination of these properties in a single CFPS results in a powerful tool that can be incorporated into current clinical practices to assess facial dysmorphism.

Abstract

Abstract
With the ever-growing complexity of deep learning models for face recognition, it becomes hard to deploy these systems in real life. Researchers have two options: 1) use smaller models; 2) compress their current models. Since the usage of smaller models might lead to concerning biases, compression gains relevance. However, compressing might be also responsible for an increase in the bias of the final model. We investigate the overall performance, the performance on each ethnicity subgroup and the racial bias of a State-of-the-Art quantization approach when used with synthetic and real data. This analysis provides a few more details on potential benefits of performing quantization with synthetic data, for instance, the reduction of biases on the majority of test scenarios. We tested five distinct architectures and three different training datasets. The models were evaluated on a fourth dataset which was collected to infer and compare the performance of face recognition models on different ethnicity.