In the paper "Learning both Weights and Connections for Efficient Neural Networks", it shrink the size of a neural networks according to a criteria. However, I found that all of the implementation of this paper does not physically shrink the size of the model. They are just zeroing the weight of some neurons to get the same result. But physically the size of the model and the inference time do not change. I'm wondering how could I physically do this?
Asked
Active
Viewed 207 times
