Google Professional Machine Learning Engineer Question 215
Single answerGoogle Cloud PlatformYou are working on a machine learning project to classify medical images for diagnosing diseases. The dataset consists of 1 million high-resolution labeled images. The team is concerned about training time and resource consumption. You also need to implement transfer learning to leverage pre-trained models. Which ML framework and model architecture combination should you choose?
- A
Use TensorFlow with a pre-trained EfficientNet model for transfer learning.
- B
Use Scikit-learn with a Random Forest classifier.
- C
Use TensorFlow with a custom convolutional neural network (CNN) built from scratch.
- D
Use PyTorch with a pre-trained LSTM model for transfer learning.
Show answer and explanation
Correct answer: A
Explanation
The problem involves classifying a large dataset of high-resolution images, which is best suited for deep learning models like CNNs. Transfer learning can save time and resources by using pre-trained models. TensorFlow supports pre-trained EfficientNet models that are optimized for image classification, making it the most suitable choice in this scenario.
- A. Correct.
EfficientNet is a state-of-the-art model for image classification tasks, and TensorFlow has robust support for transfer learning using pre-trained models like EfficientNet. This choice balances resource efficiency with accuracy and leverages pre-trained weights to speed up training.
- B. Incorrect.
Scikit-learn is not designed for deep learning or handling high-dimensional data like high-resolution images. A Random Forest classifier would not be suitable for this task.
- C. Incorrect.
Building a custom CNN from scratch would require significant time and computational resources. Additionally, it would not leverage the advantages of transfer learning, making it less efficient for large datasets.
- D. Incorrect.
LSTM models are designed for sequential data like text or time series, not for image classification tasks. Using PyTorch with an LSTM for this use case would be inappropriate.