Towards a Gender Inclusive Neural Network for Automatic Gender Recognition

Today's data-driven systems and official statistics often oversimplify the concept of gender, reducing it to binary data, which carries far-reaching implications for policy development and equitable access to services. This simplification tends to result in misclassification and discrimination against individuals who identify as gender-nonconforming. The proposed research aims to develop new, more equitable approaches that can effectively circumvent discrimination based on gender identity. Within this research framework, the primary emphasis lies in addressing the problem of underrepresentation and, in some instances, the complete absence of gender-nonconforming individuals in data collection efforts. The work presented herein represents an initial endeavor to design an equitable neural network capable of accurately identifying gender within a multiclass context, including individuals whose gender identity transcends the binary spectrum. To achieve this objective, a comprehensive comparative analysis was conducted on several fine-tuned neural network models. The aim was to acquire a profound understanding of the pivotal distinguishing features in gender identification classification and to depict the limitations of current methodologies through the application of explainable AI techniques. Different data collection techniques were tested in order to built a profitable dataset in a field where a suitable date collection is missing. The goal has been partially achieved, but the problem of invisibility is reflected in the availability of available and labeled data. The fields of use that have been envisioned are those of inclusive communication. The creation of societal maps, useful for gaining representation, and the creation of a fairness score for cluster of images were the two main ideas for a fair use of the technology proposed. The initial findings indicate promise, showcasing the efficacy of a fine-tuned EfficientNetB0 model in precisely categorizing images of individuals according to their self-reported gender. However, there exists skepticism regarding its applicability in real-world scenarios due to the limited amount of data currently available concerning non-binary individuals.