Model pruning

noun

Model pruning is a process that trims down an AI model by removing less important neurons, weights, or layers, making it more efficient without significantly hurting performance.

It works by analyzing which parts of the model contribute the least to predictions and deleting them, often using techniques like magnitude-based pruning or structured pruning. This reduces the model’s size, speeds up inference, and lowers computational costs.

For developers, pruning means they can deploy AI models on cheaper hardware or mobile devices, but it may require retraining to maintain accuracy. For example, a large language model like LLaMA-2 (70B parameters) can be pruned to 30B parameters, reducing inference time by 40% while losing only 2% accuracy.

A company uses model pruning to shrink a chatbot's AI model, allowing it to run smoothly on a smartphone instead of a cloud server.

See also: Neural network · Inférence

Découvrez nos autres services

ElpisIA, c’est tout un univers — explorez nos autres services.