Floating-Point Mantissa Rounding for Unbiased Precision Reduction

The patent applies preliminary action by performing stochastic rounding during the conversion process from single-precision to half-precision format. A random number is generated and added to the mantissa before truncation, which preliminarily distributes the rounding error statistically rather than deterministically. This preliminary randomization prevents systematic bias accumulation in subsequent neural network operations while achieving the desired memory compression.

The patent applies parameter changes by modifying the rounding behavior from deterministic to stochastic. A random parameter (random number) is introduced into the rounding process, changing the nature of the operation from fixed to probabilistic. This parameter change ensures that rounding errors follow a statistical distribution with zero mean, preventing systematic bias while maintaining computational efficiency through hardware-friendly random number generation.

The patent applies local quality by using different precision formats for different parts of the neural network system. Critical operations that require high accuracy maintain single-precision format, while storage and less critical computations use half-precision format. The stochastic rounding conversion enables this heterogeneous precision approach, allowing the system to optimize memory bandwidth efficiency in data storage and transfer while preserving computational accuracy where needed.