A manufacturer has 8 defect classes and provided ~100,000 images per class. During training, the model achieves 90% training accuracy but only 80% validation accuracy, while human-level performance is ~90%. What should you investigate to address this gap?

The significant gap between training accuracy (90%) and validation accuracy (80%), coupled with human-level performance at 90%, strongly suggests the model is overfitting. Overfitting occurs when a model learns the training data too well, including noise, and fails to generalize to unseen data. Regularization techniques (e.g., L1/L2 regularization, dropout, data augmentation) penalize complex models or introduce noise during training, forcing the model to learn more robust and generalizable features. Training for a longer time would likely exacerbate overfitting. Increasing network capacity would also increase the model's ability to memorize the training data, making overfitting worse. While a different optimizer might slightly improve convergence, it won't fundamentally address the generalization issue caused by overfitting.