Model Compression

Model compression (model compression) is a whole of techniques that carries the purpose of protecting the performance as much as possible while reducing the size of an artificial intelligence model, memory usage and calculation cost. Three main approaches stand out: quantitative (quantization) allows memory and calculation savings by representing the weight of the model normally with lower precision formats instead of high precision number formats used; such as (8-bit or 4-bit full numbers), budama (pruning) is very few contributors to the performance of the model and watching the network by removing neurons; information distillation (knowledge dispersion) is a great and powerful "student" model information, a much smaller "student" model, which transfers the teacher by mimicking.

The importance of model compression comes from the fact that the most powerful models often require a huge memory and calculation source, but the real world applications often have to work in restricted environments such as mobile phones, IoT devices or limited cloud resources. A compressed model makes it possible for many scenarios from voice assistants that can work offline on smartphones, to low lag real-time applications; today the open source is thanks to the ripening of these techniques, which can be operated in consumer hardware of quantitative versions.