A low power neural network training processor with 8-bit floating point with a shared exponent bias and fused multiply add trees | IEEE Conference Publication | IEEE Xplore