A Machine Learning Specialist is developing a mode that classifies defective parts from a manufacturing process into one of eight defect types. The training data consists of 100,000 images per…

A longer training time would only be effective if the model had not converged; given the training accuracy is already 90% and close to validation, simply training longer is unlikely to bridge the significant gap to 97% human performance.

Data augmentation synthetically expands the training dataset by creating modified versions of existing images, which helps the model learn more robust and generalized features, effectively addressing underfitting and improving performance towards human-level accuracy.

Even with a large initial dataset, acquiring more diverse or representative training data can provide the model with additional examples to learn from, helping to reduce bias and close the gap between current model performance and human-level accuracy.

While a different optimizer might slightly improve convergence, it is less likely to resolve the fundamental issue of high bias/underfitting that suggests the model is not capturing the underlying complexity of the data.

L2 regularization is a technique used to mitigate overfitting by penalizing model complexity. Applying it to an underfitting model would further restrict its learning capacity and likely worsen its performance.