3 ms·
Is integer training insufficient for deep nets? (big implications for DL chips)
- deepnotderp 10y agoBasically, they show that 40-bits(!!) are necessary for GoogLeNet inference whereas only 16 are necessary with floats. Additionally they show that XNOR-net for GoogLeNet doesn't converge.