Supervised learning models in deep learning rely heavily on large, high-quality labelled datasets, but the presence of noisy labels can severely hinder model performance and generalisation. Traditional methods treat all training examples equally, which can lead to overfitting and poor robustness when noisy labels are present. IF addresses this problem in the presence of noisy labels in supervised learning by estimating the impact of each training example on model predictions. By identifying and reducing the weights of noisy or mislabeled examples, IF-based adjustment of training weights can improve model robustness and generalisation without manual inspection. This paper shows how this approach can improve the accuracy and reliability of models in high-risk domains such as healthcare and social media. Experimental results show that IF is effective in improving model performance by effectively utilizing the training data.