Objective: The objective of the study is to investigate how biases in emotion data affect the inference process of artificial intelligence (AI) models. The first study visually examines the inference process of two distinct models trained on positively biased and negatively biased data, respectively, to observe differences. The second study investigates captions generated by image captioning models trained on four datasets with different ratios of positive and negative biases in emotion.
Background: As AI development accelerates with diverse and large-scale data, concerns about biases raising erroneous learning and reduced prediction accuracy have emerged. Data biases cause unfair outcomes related to race, gender, or ideology.
Method: The first study explores the influence of emotion data bias on a CNN model"s inference process, revealing concentration patterns for positively and negatively biased emotion data using LayerCAM. The second quantitatively analyzes bias influence on outcomes using an image captioning model trained on differently biased datasets. We quantitatively analyze the relationship between the generated captions and the datasets in terms of the Sentiment Intensity Analysis (SIA) values.
Results: Divergent inference processes were observed in the first study, showing that the focused regions of the models were different by different biases in emotion data trained. The second study demonstrated that as the bias ratio shifted, the model-generated captions changed accordingly, with a tendency to produce more positive captions with decreased positive and increased negative biases.
Conclusion: This study underscores the critical impact of emotion data bias on AI model inference and outcomes. Recognizing and mitigating biases in emotion data is imperative for developing fair and effective AI models, particularly in applications involving human-AI interaction. In the context of AI integration in services, addressing emotion data bias is crucial for responsible AI deployment, especially in human-AI interaction.