Bias Mitigation in Deep Learning: A Survey of Modern Techniques

Ensuring fairness in AI systems is crucial, particularly in medical imaging, natural language processing (NLP), and recommender systems. This survey reviews modern bias mitigation techniques based on a systematic review of 100 studies, narrowing down to fewer than 20 with significant contributions to AI fairness. The techniques are categorized into data-centric, algorithmic, and hybrid approaches, including strategies such as data augmentation, feature selection, fairness-aware metrics, and attention mechanisms. Notable methods include AttEN for dermatological prediction, VERB for debiasing word embeddings, and FairIF for fairness improvements via influence functions. Domain-specific solutions address bias in coronary disease diagnosis and toxic comment detection, while frameworks like EBRank enhance fairness in ranking systems. This paper evaluates these techniques in terms of effectiveness, computational efficiency, and applicability, highlighting the role of hybrid approaches in advancing equitable AI development.