Directing DNNs Attention for Facial Attribution Classification using Gradient-weighted Class Activation Mapping

DNNs can learn the wrong features: The paper demonstrates that deep neural networks trained on biased datasets may rely on incorrect visual features for classification decisions, which undermines their ability to work reliably when applied to new situations or datasets.

Co-occurrence bias is a real problem: The research identifies "co-occurrence bias"—where certain features appear together in training data—as a significant issue that can cause AI models to make decisions based on spurious correlations rather than the actual attributes they're supposed to identify.

Human guidance can improve AI decision-making: The proposed method allows users to manually direct the model's attention to specific facial regions, suggesting that incorporating human oversight into the training process can help correct biased learning patterns and improve model reliability.

Visualization tools reveal what AI is actually looking at: By using Grad-CAM (a technique that shows which image regions the model focuses on), researchers can identify when models are attending to the wrong features, making AI decision-making more transparent and auditable.

Improved transferability benefits real-world deployment: By mitigating co-occurrence bias through directed attention, the fine-tuned models show better performance when applied to new datasets, which is crucial for responsible AI systems that need to work reliably across different populations and contexts.