Google Professional Machine Learning Engineer Question 349
Select 3Google Cloud PlatformYou are designing a machine learning model to classify customer reviews as positive or negative. The model will be deployed in production to handle real-time predictions. During initial testing, you notice that the model has high latency and consumes excessive memory. Which of the following strategies can you use to optimize the model for production without significantly compromising accuracy?
- A
Use model quantization to reduce the precision of weights and activations.
- B
Increase the size of the input features to ensure more context is captured.
- C
Prune unnecessary neurons or layers from the model.
- D
Use a smaller, pre-trained model architecture such as MobileNet.
- E
Enable batch predictions for real-time serving.
Show answer and explanation
Correct answers: A, C, D
Explanation
Optimizing machine learning models for production involves simplifying the model to reduce latency and memory usage while maintaining acceptable accuracy. Techniques like quantization, pruning, and using smaller pre-trained architectures can achieve these goals effectively. Increasing input features or enabling batch predictions would not address the performance issues in real-time serving scenarios.
- A. Correct.
Model quantization helps reduce the memory footprint and computational requirements by lowering the precision of weights (e.g., from 32-bit floating point to 8-bit integers). This is a common technique for optimizing models for production.
- B. Incorrect.
Increasing the size of input features may improve accuracy but will increase computational complexity, latency, and memory usage. This does not align with the goal of optimizing for production.
- C. Correct.
Pruning removes unnecessary neurons or layers from the model, reducing its size and improving inference efficiency without significantly impacting accuracy.
- D. Correct.
Using a smaller, pre-trained model architecture like MobileNet is an effective way to simplify the model and make it more suitable for production environments with limited resources.
- E. Incorrect.
Batch predictions are used to handle large volumes of offline predictions rather than real-time predictions. They are not suitable for optimizing a model for real-time serving.