MLA-C01 Practice Question 64: JumpStart Transfer Learning
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 7 · question 4 of 10
An online retailer needs to sort product photos into 12 in-house categories. It has 1,500 labeled photos in total, and its data scientists need a model they can tune with their own hyperparameters and host on the company's existing SageMaker AI endpoints next to other models. The team wants good accuracy quickly and with little training compute.
Which approach should the team take?
Show answer and explanation
Answer:
Fine-tuning trains a pretrained model on a new dataset without starting from scratch. This is transfer learning, and it can produce accurate models from smaller datasets with less training time. A JumpStart image classification model has already learned general visual features from a large dataset, so 1,500 labeled photos are enough to adapt it to 12 new categories. The team keeps control of the hyperparameters and deploys the result to a SageMaker AI endpoint like any other model (option C).
- A is wrong because training a CNN from random weights on 1,500 images tends to overfit and needs far more data, compute, and time to reach the accuracy that fine-tuning gets quickly. - B is wrong because DetectLabels returns generic labels that were never trained on this retailer's categories. Hand-mapping them is brittle, and it is not a model the team can tune or host on its endpoints. - D is wrong because k-means is unsupervised and ignores the labels the team already has. There is no guarantee that the 12 clusters line up with the 12 business categories.
Reference: Fine-Tune a Model - SageMaker JumpStart