A company's machine learning (ML) specialist is building a computer vision model to classify 10 different traffic signs. The company has stored 100 images of each class in Amazon S3, and the company…

Overfitting is often caused by insufficient training data, so using Amazon SageMaker Ground Truth to label the 10,000 unlabeled images significantly expands the training dataset. This provides more diverse examples for the model, helping it generalize better and reducing overfitting.

Transforming images to grayscale reduces color information, which might be crucial for distinguishing traffic signs, and is not a primary or universally effective method for addressing overfitting.

Data augmentation, such as rotating and translating existing labeled images, artificially increases the size and diversity of the training dataset without requiring new data collection. This technique helps the model learn more robust features and become less sensitive to minor variations, directly combating overfitting.

Replacing the activation of the last layer with a sigmoid is appropriate for binary classification; for 10 classes, a softmax activation function is required, making this change functionally incorrect for the problem type.

While k-NN could be used for pseudo-labeling, Amazon SageMaker Ground Truth offers human-in-the-loop labeling, which provides higher quality and more reliable labels crucial for effectively addressing overfitting with new data.