This paper investigates the impact of hard-negative contrastive learning on facial emotion recognition (FER) systems, particularly in the context of noisy and real-world datasets. The authors explore the effectiveness of latent feature representations derived from contrastive learning techniques, focusing on both unsupervised (UCL) and supervised (SCL) methods. They evaluate these techniques on four benchmark datasets: FER2013, FERPlus, RAF-DB, and AffectNet. The study highlights that the choice of feature representations significantly influences the performance of FER systems. The proposed approach, which combines contrastive learning with a simple supervised classification framework, demonstrates superior performance compared to state-of-the-art methods, even with smaller batch sizes and fewer epochs. The results suggest that hard-negative contrastive learning can enhance the robustness and accuracy of FER systems in noisy and real-world scenarios.This paper investigates the impact of hard-negative contrastive learning on facial emotion recognition (FER) systems, particularly in the context of noisy and real-world datasets. The authors explore the effectiveness of latent feature representations derived from contrastive learning techniques, focusing on both unsupervised (UCL) and supervised (SCL) methods. They evaluate these techniques on four benchmark datasets: FER2013, FERPlus, RAF-DB, and AffectNet. The study highlights that the choice of feature representations significantly influences the performance of FER systems. The proposed approach, which combines contrastive learning with a simple supervised classification framework, demonstrates superior performance compared to state-of-the-art methods, even with smaller batch sizes and fewer epochs. The results suggest that hard-negative contrastive learning can enhance the robustness and accuracy of FER systems in noisy and real-world scenarios.