Does Hard-Negative Contrastive Learning Improve Facial Emotion Recognition? This paper investigates the impact of hard-negative contrastive learning on facial emotion recognition (FER) in real-world scenarios with noisy labels. The authors propose a framework that learns latent feature representations through contrastive learning without requiring preprocessing steps. The extracted features are then used in a perceptron network for supervision. The main contributions include studying instance-level feature representation learning in FER, exploring the effect of latent representations from noisy FER datasets on fundamental supervised frameworks, and conducting experiments on noisy datasets that demonstrate superior performance compared to benchmark contrastive methods. The study evaluates the proposed framework on four benchmark datasets: FER2013, FERPlus, RAF-DB, and AffectNet. The results show that the choice of feature representations significantly impacts FER performance. The paper also discusses the benefits of using sophisticated feature representations within fundamental architectural frameworks. The authors compare different contrastive learning approaches, including unsupervised contrastive learning (UCL), supervised contrastive learning (SCL), and their variants with hard-negative samples. The results indicate that the proposed hard-negative contrastive learning approach outperforms existing methods, particularly in noisy environments. The study also explores the use of generalized cross-entropy loss for training deep neural networks with noisy labels, which is more robust than standard cross-entropy loss. The authors demonstrate that their proposed method achieves higher accuracy on the RAF-DB dataset compared to existing methods. The paper concludes that the combination of pre-trained representation learning and simple learning frameworks can lead to remarkable performances in FER. The findings suggest that the proposed method can serve as a new baseline for learning in other real-world datasets and classification domains.Does Hard-Negative Contrastive Learning Improve Facial Emotion Recognition? This paper investigates the impact of hard-negative contrastive learning on facial emotion recognition (FER) in real-world scenarios with noisy labels. The authors propose a framework that learns latent feature representations through contrastive learning without requiring preprocessing steps. The extracted features are then used in a perceptron network for supervision. The main contributions include studying instance-level feature representation learning in FER, exploring the effect of latent representations from noisy FER datasets on fundamental supervised frameworks, and conducting experiments on noisy datasets that demonstrate superior performance compared to benchmark contrastive methods. The study evaluates the proposed framework on four benchmark datasets: FER2013, FERPlus, RAF-DB, and AffectNet. The results show that the choice of feature representations significantly impacts FER performance. The paper also discusses the benefits of using sophisticated feature representations within fundamental architectural frameworks. The authors compare different contrastive learning approaches, including unsupervised contrastive learning (UCL), supervised contrastive learning (SCL), and their variants with hard-negative samples. The results indicate that the proposed hard-negative contrastive learning approach outperforms existing methods, particularly in noisy environments. The study also explores the use of generalized cross-entropy loss for training deep neural networks with noisy labels, which is more robust than standard cross-entropy loss. The authors demonstrate that their proposed method achieves higher accuracy on the RAF-DB dataset compared to existing methods. The paper concludes that the combination of pre-trained representation learning and simple learning frameworks can lead to remarkable performances in FER. The findings suggest that the proposed method can serve as a new baseline for learning in other real-world datasets and classification domains.