Method and system for elastic precision enhancement using dynamic shifting in neural networks

A long bit-width multiplier for calculating a multiplication of a first fixed point number and a second fixed point number, the first fixed point number being associated with a pixel of an input feature map corresponding to one of a plurality of channels with respect to a layer of a neural network having a plurality of layers, and the second fixed point number being associated with a filter corresponding to the one of the plurality of channels with respect to the layer of the neural network, the long bit-width multiplier comprising:

• a first converter configured to convert the first fixed point number to a first sign, a first mantissa, and a first exponent, wherein at least one of a bit width of the first sign, a bit width of the first mantissa and a bit width of the first exponent is dynamically configured based on one or more factors selected from a group consisting of a first relative position of the layer in the neural network, a second relative position of the pixel in the input feature map, and the one of the plurality of channels;
• and
• a restoration circuit configured to calculate and output the multiplication of the first fixed point number and the second fixed point number based on the first sign, the first mantissa, the first exponent, and the second fixed point number.