Fig. 3From: Neural network training with limited precision and asymmetric exponent32-bit floating-point number representationBack to article page