Google

Pruning is a technique for compressing BERT by strategically removing parts of its Transformer network. This can be implemented in several ways, such as eliminating entire layers, removing a certain percentage of network parameters, or discarding specific attention heads. These actions can significantly speed up model inference, often without a major decrease in performance.

Pruning for BERT Compression

A development team needs to deploy a large, pre-trained Transformer-based language model on a mobile device with limited processing power. Their primary goal is to significantly reduce inference time while minimizing the loss of accuracy on a sentiment analysis task. They are considering two different strategies for compressing the model. Evaluate the two proposals below and justify which one is more likely to achieve the team's goal.

Evaluating Model Compression Strategies

A machine learning team is exploring different methods to reduce the size and inference time of a large language model based on the Transformer architecture. Match each pruning technique with its most likely description or primary impact.

A development team needs to accelerate the inference speed of a large, pre-trained language model for a task requiring a deep understanding of long-range dependencies and complex sentence structures. Which of the following strategies for reducing the model's size is most likely to severely degrade performance on this specific task?

Learn Before

Related