When presenting this slide, Kayvon pointed out the last column, which shows the percentage of the original data that the compressed version of the model is. The fact that the percentages range from roughly 1% to 30% show how incredibly valuable these software-based compression approaches are. With barely any difference in accuracy, these models are shrunk WAY down, which is good news for the future of DNN application at scale, or on mobile devices for example.
When presenting this slide, Kayvon pointed out the last column, which shows the percentage of the original data that the compressed version of the model is. The fact that the percentages range from roughly 1% to 30% show how incredibly valuable these software-based compression approaches are. With barely any difference in accuracy, these models are shrunk WAY down, which is good news for the future of DNN application at scale, or on mobile devices for example.