MLPER-10: Detect performance issues when using transfer learning
Monitor and ensure that the inherited prediction weights from a transferred model yield the desired results. This approach helps minimize the risk of weak learning and incorrect outputs using pre-trained models.
Implementation plan
-
Use Amazon SageMaker Debugger - Transfer learning is a machine learning technique where a model pre-trained on one task is fine-tuned on a new task. When using the transfer learning approach, use Amazon SageMaker Debugger to detect hidden problems that might have serious consequences. Examine model predictions to see what mistakes were made. Validate the robustness of your model, and consider how much of this robustness is from the inherited capabilities. Validate input and preprocesses to the model for realistic expectations.
Blogs
