Google Professional Machine Learning Engineer Question 218
Single answerGoogle Cloud PlatformYou are tasked with building a machine learning model for a large-scale image classification problem on Google Cloud. The dataset consists of millions of high-resolution images, and the solution must be optimized for both training speed and prediction latency. You have decided to use one of Google Cloud's managed services for model training. Which combination of ML framework and model architecture should you choose to meet these requirements?
- A
TensorFlow with a pre-trained EfficientNet model fine-tuned on your dataset
- B
PyTorch with a custom-built convolutional neural network (CNN) trained from scratch
- C
Scikit-learn with a random forest classifier applied to downsampled image features
- D
TensorFlow with a simple fully connected neural network trained from scratch
Show answer and explanation
Correct answer: A
Explanation
For large-scale image classification, using TensorFlow with a pre-trained EfficientNet model is the most appropriate choice. EfficientNet is a state-of-the-art architecture that balances high accuracy with computational efficiency, making it suitable for large datasets. TensorFlow integrates seamlessly with Google Cloud's managed services, enabling efficient training and deployment while meeting the requirements for training speed and prediction latency.
- A. Correct.
TensorFlow with a pre-trained EfficientNet model is an optimal choice for large-scale image classification. EfficientNet is designed for high accuracy with lower computational requirements, and TensorFlow integrates well with Google Cloud's managed services, enabling fast training and low-latency predictions.
- B. Incorrect.
PyTorch is a flexible and popular framework, but training a custom CNN from scratch on millions of high-resolution images is computationally expensive and time-consuming. This approach is not ideal given the focus on training speed and prediction latency.
- C. Incorrect.
Scikit-learn is not well-suited for large-scale image classification problems, especially those involving millions of high-resolution images. Random forest classifiers do not perform well on such tasks, and downsampling image features would significantly reduce model performance.
- D. Incorrect.
A fully connected neural network trained from scratch is not suitable for image classification tasks, as it lacks the spatial feature extraction capabilities of convolutional architectures. It would also require significantly more resources and time to train on high-resolution images.