Do we really need model compression?

Yes, there is a humongous need of model compression as neural networks are often over-parameterized, and there are many parameter redundancy, which is not conducive to mobile deployment. Computing platforms with large computing power are often very expensive, and model compression can directly save hardware costs. Model compression extracts the "simple" model from the large model by eliminating redundancy, so that the memory and time efficiency are closer to the ideal appropriate parameterized model.

Share it on social networks: Tweet Share